LM-Kit Model Catalog

70 curated models from 19 vendors, spanning chat, reasoning, vision, OCR, speech-to-text, embeddings, and reranking. Each entry ships with its SHA-256 checksum, downloads on first use, and runs entirely on your hardware.

Load any model by its ID; LM-Kit resolves, downloads, and caches it automatically:

var model = LM.LoadFromModelID("qwen3.5:9b");

Click any row for the full specification, checksum, loading snippet, and direct download. Embedding models show their MTEB quality score.

All models in the LM-Kit catalog with their capabilities and specifications
Model	Capabilities	Params	Context	Download	License
BAAI bge m3 `bge-m3` MTEB 59.6	embeddings	567 M	8K	417 MB	mit
A unified, multilingual embedding model that delivers dense, sparse, and multi-vector retrieval on texts from short queries up to 8,192-token documents in over 100 languages. Architecturebert Parameters566,703,104 Context window8,192 tokens Quantization4-bit FormatGGUF Download size417.5 MB Licensemit MTEB score59.6 · leaderboard SHA-256`e251234fcb7d050991a6be491952f485bf5c641dd10c3272dc1301fd281ad50f` `var model = LM.LoadFromModelID("bge-m3");` Download bge-m3-Q4_K_M.gguf417 MB
BAAI bge m3 reranker v2 `bge-m3-reranker`	reranking	568 M	8K	418 MB	apache-2.0
A unified, multilingual reranker that ingests query–document pairs and directly produces sigmoid-normalized relevance scores across over 100 languages. Architecturebert Parameters567,753,729 Context window8,192 tokens Quantization4-bit FormatGGUF Download size418.07 MB Licenseapache-2.0 SHA-256`ce947cece730cbf7d836da8c5490a9987ef0f919014b9275e7ce9aa12d96e6d9` `var model = LM.LoadFromModelID("bge-m3-reranker");` Download Bge-M3-568M-Q4_K_M.gguf418 MB
BAAI bge small en v1.5 `bge-small`	embeddings	33 M	512	64 MB	mit
An efficient, CPU-friendly English embedding model (BAAI General Embedding) designed for lightweight applications. Architecturebert Parameters33,212,160 Context window512 tokens Quantization16-bit FormatGGUF Download size64.45 MB Licensemit SHA-256`cd5790da23df71e7e20fe20bb523bd4586a533a4ee813cc562e32b37929141c1` `var model = LM.LoadFromModelID("bge-small");` Download bge-small-en-v1.5-f16.gguf64 MB
DeepSeek Coder V2 Lite `deepseek-coder-v2:16b`	code completion	15.7 B	160K	9.7 GB	deepseek
An open-source mixture-of-experts code model tailored for code completion tasks. Early evaluations indicated competitive performance relative to leading code models. Architecturedeepseek2 Parameters15,706,484,224 Context window163,840 tokens Quantization4-bit FormatGGUF Download size9.65 GB Licensedeepseek SHA-256`ac398e8c1c670d3c362d3c1182614916bab7c364708ec073fcf947f6802d509e` `var model = LM.LoadFromModelID("deepseek-coder-v2:16b");` Download DeepSeek-Coder-2-Lite-15.7B-Instruct-Q4_K_M.gguf9.7 GB
DeepSeek R1 Distill Llama `deepseek-r1:8b`	text generation, chat, code completion, math	8.03 B	128K	4.6 GB	mit
DeepSeek-R1 enhances its predecessor by integrating cold-start data to overcome repetition and readability issues, achieving state-of-the-art performance in math, code, and reasoning tasks, with all models open-sourced. Architecturellama Parameters8,030,261,312 Context window131,072 tokens Quantization4-bit FormatGGUF Download size4.58 GB Licensemit SHA-256`596fce705423e44831fe63367a30ccc7b36921c1bfdd4b9dfde85a5aa97ac2ef` `var model = LM.LoadFromModelID("deepseek-r1:8b");` Download DeepSeek-R1-Distill-Llama-8B-Q4_K_M.gguf4.6 GB
Mistral Devstral Small 2 `devstral-small2`	text generation, chat, code completion, vision, tool calling	24 B	384K	14.2 GB	apache-2.0
Agentic LLM for software engineering tasks. Excels at using tools to explore codebases, editing multiple files, and powering software engineering agents. Features vision capabilities for analyzing images alongside code understanding. Architecturemistral3 Parameters24,011,366,400 Context window393,216 tokens Quantization4-bit FormatLMK Download size14.17 GB Licenseapache-2.0 SHA-256`6d391403096e1b386e746183d1714ad9579dadf89e36d31c52c90a7de579bf97` `var model = LM.LoadFromModelID("devstral-small2");` Download devstral-small-2-24b-instruct-Q4_K_M.lmk14.2 GB
Google Gemma Embedding 300M `embeddinggemma-300m` MTEB 61.15	embeddings	303 M	2K	225 MB	gemma
EmbeddingGemma 300M is an open, state-of-the-art-for-its-size embedding model from Google DeepMind (Gemma 3, T5Gemma-initialized). It produces 768-dimensional text vectors for search/retrieval, classification, clustering, and semantic similarity across 100+ languages, and supports Matryoshka Representation Learning (truncate to 512/256/128 with re-normalization). Optimized for on-device/CPU deployment. Architecturegemma-embedding Parameters302,863,104 Context window2,048 tokens Quantization4-bit FormatGGUF Download size225.39 MB Licensegemma MTEB score61.15 · leaderboard SHA-256`3d55e7fe66eb4c7b2d01b4fbd30c00dc7a101bd6c9f724a6e7e5cfaa87968420` `var model = LM.LoadFromModelID("embeddinggemma-300m");` Download embeddinggemma-300M-Q4_K_M.gguf225 MB
TII Falcon H1R `falcon-h1r:7b`	text generation, chat, code completion, math, reasoning, tool calling	7.59 B	256K	4.3 GB	falcon-llm-license
Reasoning-specialized hybrid (Transformers + Mamba2) Falcon-H1R model, trained with long reasoning traces and GRPO RL. Strong on math, coding, instruction following, and general logic. Architecturefalcon-h1 Parameters7,585,648,736 Context window262,144 tokens Quantization4-bit FormatGGUF Download size4.28 GB Licensefalcon-llm-license SHA-256`0bf6a7af23cb42d5b69ea486a4bb67a76cdb7f51a56419cca72cee8fd19f010c` `var model = LM.LoadFromModelID("falcon-h1r:7b");` Download Falcon-H1R-7B-Q4_K_M.gguf4.3 GB
TII Falcon 3 Instruct `falcon3:3b`	text generation, chat, code completion, math	3.23 B	32K	1.9 GB	falcon-llm-license
Designed for multilingual tasks including chat, text generation, and code completion, supporting extended context lengths. Architecturellama Parameters3,227,655,168 Context window32,768 tokens Quantization4-bit FormatGGUF Download size1.87 GB Licensefalcon-llm-license SHA-256`81c6b52d221c2f0eea3db172fc74de28534f2fd15f198ecbfcc55577d20cbf8a` `var model = LM.LoadFromModelID("falcon3:3b");` Download Falcon3-3B-Instruct-q4_k_m.gguf1.9 GB
TII Falcon 3 Instruct `falcon3:7b`	text generation, chat, code completion, math	7.62 B	32K	4.3 GB	falcon-llm-license
Offers robust performance across chat, text generation, and mathematical reasoning tasks with extended context support. Architecturellama Parameters7,615,616,512 Context window32,768 tokens Quantization4-bit FormatGGUF Download size4.26 GB Licensefalcon-llm-license SHA-256`4ce1da546d76e04ce77eb076556eb25e1096faf6155ee429245e4bfa3f5ddf5d` `var model = LM.LoadFromModelID("falcon3:7b");` Download Falcon-3-7.6B-Instruct-Q4_K_M.gguf4.3 GB
TII Falcon 3 Instruct `falcon3:10b`	text generation, chat, code completion, math	10.3 B	32K	5.9 GB	falcon-llm-license
A larger variant tailored for multilingual dialogue, code completion, and complex reasoning tasks with extended context support. Architecturellama Parameters10,305,653,760 Context window32,768 tokens Quantization4-bit FormatGGUF Download size5.86 GB Licensefalcon-llm-license SHA-256`a0c0edbd35019ff26d972a0373b25b4c8d72315395a3b6036aca5e6bafa3d819` `var model = LM.LoadFromModelID("falcon3:10b");` Download Falcon3-10B-Instruct-q4_k_m.gguf5.9 GB
Google Gemma 2legacy `gemma2:2b` superseded by `gemma4:e2b`	text generation, chat	2.61 B	8K	1.6 GB	gemma
Legacy model. Superseded by `gemma4:e2b`; it remains available for compatibility, but new projects should use the replacement. A lightweight decoder-only model from Google, available in both pre-trained and instruction-tuned variants for text-to-text tasks. Architecturegemma2 Parameters2,614,341,888 Context window8,192 tokens Quantization4-bit FormatGGUF Download size1.59 GB Licensegemma SHA-256`362d09c1496e035ecf0737d8fe03e8e607c61e57e16b22cedd158525f6721e06` `var model = LM.LoadFromModelID("gemma2:2b");` Download gemma-2-2B-Q4_K_M.gguf1.6 GB
Google Gemma 2legacy `gemma2:9b` superseded by `gemma4:e4b`	text generation, chat	9.24 B	8K	5.4 GB	gemma
Legacy model. Superseded by `gemma4:e4b`; it remains available for compatibility, but new projects should use the replacement. A decoder-only text-to-text model from Google, offering competitive performance in both pre-trained and instruction-tuned configurations. Architecturegemma2 Parameters9,241,705,984 Context window8,192 tokens Quantization4-bit FormatGGUF Download size5.37 GB Licensegemma SHA-256`b6059a960d2f4f881630f1e795b40f7e09e5e12d3a6b1900474d6108ea880afd` `var model = LM.LoadFromModelID("gemma2:9b");` Download gemma-2-9B-Q4_K_M.gguf5.4 GB
Google Gemma 2legacy `gemma2:27b` superseded by `gemma4:26b-a4b`	text generation, chat	27.2 B	8K	15.5 GB	gemma
Legacy model. Superseded by `gemma4:26b-a4b`; it remains available for compatibility, but new projects should use the replacement. A larger variant in the Gemma 2 family, optimized for text generation and instruction following with open weights provided. Architecturegemma2 Parameters27,227,128,320 Context window8,192 tokens Quantization4-bit FormatGGUF Download size15.5 GB Licensegemma SHA-256`bb4b276745da743d550720dc2e6c847498eef45e7b82a4d5a73ef6636f78027a` `var model = LM.LoadFromModelID("gemma2:27b");` Download gemma-4-26B-A4B_q4_0-it.lmk15.5 GB
Google Gemma 3legacy `gemma3:1b` superseded by `gemma4:e2b`	text generation, chat	1000 M	32K	769 MB	gemma
Legacy model. Superseded by `gemma4:e2b`; it remains available for compatibility, but new projects should use the replacement. Gemma is Google's lightweight, multimodal, open AI model family based on Gemini technology, supporting text and image inputs, 128K context windows, multilingual capabilities in over 140 languages, and optimized for resource-limited environments. Architecturegemma3 Parameters999,885,952 Context window32,768 tokens Quantization4-bit FormatGGUF Download size768.72 MB Licensegemma SHA-256`bacfe3de6eee9fba412d5c0415630172c2a602dae26bb353e1b20dd67194a226` `var model = LM.LoadFromModelID("gemma3:1b");` Download gemma-3-it-1B-Q4_K_M.gguf769 MB
Google Gemma 3legacy `gemma3:4b` superseded by `gemma4:e4b`	text generation, chat, code completion, math, vision	3.88 B	128K	2.9 GB	gemma
Legacy model. Superseded by `gemma4:e4b`; it remains available for compatibility, but new projects should use the replacement. Gemma is Google's lightweight, multimodal, open AI model family based on Gemini technology, supporting text and image inputs, 128K context windows, multilingual capabilities in over 140 languages, and optimized for resource-limited environments. Architecturegemma3 Parameters3,880,099,328 Context window131,072 tokens Quantization4-bit FormatGGUF Download size2.87 GB Licensegemma SHA-256`abb283e96c0abf58468a18127ce6e8b2bfc98e48f1ec618f658495c09254bdae` `var model = LM.LoadFromModelID("gemma3:4b");` Download gemma-3-4b-it-Q4_K_M.lmk2.9 GB
Google Gemma 3legacy `gemma3:12b` superseded by `gemma4:12b`	text generation, chat, code completion, math, vision	11.8 B	128K	7.4 GB	gemma
Legacy model. Superseded by `gemma4:12b`; it remains available for compatibility, but new projects should use the replacement. Gemma is Google's lightweight, multimodal, open AI model family based on Gemini technology, supporting text and image inputs, 128K context windows, multilingual capabilities in over 140 languages, and optimized for resource-limited environments. Architecturegemma3 Parameters11,765,788,416 Context window131,072 tokens Quantization4-bit FormatGGUF Download size7.35 GB Licensegemma SHA-256`d6f01cdb4369769ea87c5211a7fd865e12dbb9e2a937b43ef281a5b7e9ba2e35` `var model = LM.LoadFromModelID("gemma3:12b");` Download gemma-3-12b-it-Q4_K_M.lmk7.4 GB
Google Gemma 3legacy `gemma3:27b` superseded by `gemma4:26b-a4b`	text generation, chat, code completion, math, vision	27 B	128K	16.0 GB	gemma
Legacy model. Superseded by `gemma4:26b-a4b`; it remains available for compatibility, but new projects should use the replacement. Gemma is Google's lightweight, multimodal, open AI model family based on Gemini technology, supporting text and image inputs, 128K context windows, multilingual capabilities in over 140 languages, and optimized for resource-limited environments. Architecturegemma3 Parameters27,009,002,240 Context window131,072 tokens Quantization4-bit FormatGGUF Download size15.97 GB Licensegemma SHA-256`2d0e4382259ae2da28b9c0342e982a58eafbddad7c05bbfe6e104f2b3c165994` `var model = LM.LoadFromModelID("gemma3:27b");` Download gemma-3-27b-it-Q4_K_M.lmk16.0 GB
Google Gemma 3 270M `gemma3:270m`	text generation, chat	268 M	32K	241 MB	gemma
Gemma is Google's lightweight, multimodal, open AI model family based on Gemini technology, supporting text and image inputs, 128K context windows, multilingual capabilities in over 140 languages, and optimized for resource-limited environments. Architecturegemma3 Parameters268,098,176 Context window32,768 tokens Quantization4-bit FormatGGUF Download size241.39 MB Licensegemma SHA-256`e28b323bc75925d6edc8d3f030268830bf53c59c296d77278ac24653403d9d47` `var model = LM.LoadFromModelID("gemma3:270m");` Download gemma-3-270M-it-Q4_K_M.gguf241 MB
Google Gemma 4 `gemma4:e2b`	text generation, chat, code completion, math, vision, reasoning, tool calling	5.12 B	128K	2.9 GB	apache-2.0
Gemma 4 is Google's next generation multimodal Mixture-of-Experts AI model family (4.6B total, ~2B active). Supports text and image inputs, tool calling, 128K context windows, and multilingual capabilities in over 140 languages, optimized for efficient inference. Architecturegemma4 Parameters5,123,179,235 Context window131,072 tokens Quantization4-bit FormatLMK Download size2.87 GB Licenseapache-2.0 SHA-256`d46bb1954da3218ff0747196e6d33ad7819eb16ab1abe51a4004e86e65f5a73b` `var model = LM.LoadFromModelID("gemma4:e2b");` Download Gemma-4-E2B-It-4.6B-Q4_K_M.lmk2.9 GB
Google Gemma 4 `gemma4:e4b`	text generation, chat, code completion, math, vision, reasoning, tool calling	8.0 B	128K	4.5 GB	apache-2.0
Gemma 4 is Google's next generation multimodal Mixture-of-Experts AI model family (7.5B total, ~4B active). Supports text and image inputs, tool calling, coding, math, 128K context windows, and multilingual capabilities in over 140 languages. Architecturegemma4 Parameters7,996,157,674 Context window131,072 tokens Quantization4-bit FormatLMK Download size4.47 GB Licenseapache-2.0 SHA-256`541693f69e7d6a817019b3cb366a7b94bb82f33032d6ff62d2bc4f19ebfc9b11` `var model = LM.LoadFromModelID("gemma4:e4b");` Download Gemma-4-E4B-It-7.5B-Q4_K_M.lmk4.5 GB
Google Gemma 4 `gemma4:12b`	text generation, chat, code completion, math, vision, reasoning, tool calling	11.9 B	256K	6.8 GB	apache-2.0
Gemma 4 is Google's next generation multimodal AI model (12B dense). Features hybrid sliding and global attention with 256K context, text and image inputs, tool calling, coding, math, reasoning, and multilingual capabilities in over 140 languages. Architecturegemma4 Parameters11,907,350,576 Context window262,144 tokens Quantization4-bit FormatLMK Download size6.77 GB Licenseapache-2.0 SHA-256`1e3b78b74aed326fc5d12479c27e558b6fdca6c1598b1606d80753014abc2c67` `var model = LM.LoadFromModelID("gemma4:12b");` Download gemma-4-12B-it-Q4_K_M.lmk6.8 GB
Google Gemma 4 26B-A4B `gemma4:26b-a4b`	text generation, chat, code completion, math, vision, reasoning, tool calling	25.8 B	128K	14.3 GB	apache-2.0
Gemma 4 is Google's next generation multimodal Mixture-of-Experts AI model family (26B total, 4B active). Delivers powerful text and image understanding, tool calling, coding, math, reasoning, 128K context windows, and multilingual capabilities in over 140 languages. Architecturegemma4 Parameters25,805,936,462 Context window131,072 tokens Quantization4-bit FormatLMK Download size14.3 GB Licenseapache-2.0 SHA-256`a08cd4a0c4e3ecaf62c41673ad39342127adecf2dc6e7403b1aa68e014cc64dd` `var model = LM.LoadFromModelID("gemma4:26b-a4b");` Download gemma-4-26B-A4B_q4_0-it.lmk14.3 GB
Google Gemma 4 `gemma4:31b`	text generation, chat, code completion, math, vision, reasoning, tool calling	30.7 B	256K	17.7 GB	apache-2.0
Gemma 4 is Google's next generation multimodal AI model (31B dense). Features hybrid sliding and global attention with 256K context, text and image inputs, tool calling, coding, math, reasoning, and multilingual capabilities in over 140 languages. Architecturegemma4 Parameters30,696,952,380 Context window262,144 tokens Quantization4-bit FormatLMK Download size17.68 GB Licenseapache-2.0 SHA-256`8511cf2d1fdc2c37179829a30314e6afcd3bdabc63fcc4380a9b4ab7284d690b` `var model = LM.LoadFromModelID("gemma4:31b");` Download gemma-4-31B-it-Q4_K_M.lmk17.7 GB
Z.ai GLM-V 4.6 Flash `glm-4.6v-flash`	text generation, chat, code completion, math, vision, tool calling, ocr	10.3 B	128K	6.3 GB	mit
Lightweight GLM-V 4.6 vision-language “Flash” model optimized for low-latency local deployment. Strong at OCR and document/screenshot understanding (text, layout, charts, tables), and supports native function calling for tool-driven multimodal agents, including grounded UI/code reconstruction from images. Architectureglm4 Parameters10,292,777,472 Context window131,072 tokens Quantization4-bit FormatLMK Download size6.31 GB Licensemit SHA-256`24956fb9560c80c904cd3649562fdd11c8af65c45a3c932f785275415e34f0b8` `var model = LM.LoadFromModelID("glm-4.6v-flash");` Download glm-4.6v-flash-Q4_K_M.lmk6.3 GB
Z.ai GLM-OCR `glm-ocr`	vision, ocr	1.33 B	128K	801 MB	mit
GLM-OCR is a 0.9B vision-language model from Z.ai specialized in document parsing, OCR, and structured information extraction. Supports text, formula, table, and complex layout recognition across multiple languages. Architectureglm4 Parameters1,325,258,240 Context window131,072 tokens Quantization4-bit FormatLMK Download size801.27 MB Licensemit SHA-256`64fd1b65fac8a09b132838ddcf68b7dec4fde7eaf303cfac4e225201a48ad6bc` `var model = LM.LoadFromModelID("glm-ocr");` Download GLM-OCR.Q4_K_M.lmk801 MB
Z.ai GLM 4.7 Flash 30B-A3B `glm4.7-flash`	text generation, chat, code completion, math, reasoning, tool calling	29.9 B	198K	16.9 GB	mit
GLM-4.7-Flash is a 30B-A3B MoE model from Z.ai. As the strongest model in the 30B class, it balances performance and efficiency for lightweight deployment, excelling at agentic tasks, reasoning, coding, and math. Architecturedeepseek2 Parameters29,943,393,920 Context window202,752 tokens Quantization4-bit FormatGGUF Download size16.89 GB Licensemit SHA-256`4319acc7340fc33d084ff8e938b87f01fae43db6c608e2a7a9631ee57579fcbc` `var model = LM.LoadFromModelID("glm4.7-flash");` Download GLM-4.7-Flash-64x2.6B-Q4_K_M.gguf16.9 GB
OpenAI Gpt OSS `gptoss:20b`	text generation, chat, code completion, math, reasoning, tool calling	20.9 B	128K	11.3 GB	apache-2.0
OpenAI’s medium-sized open-weight Mixture-of-Experts model (≈21B params; ~3.6B active per token). This MXFP4 GGUF build is optimized for local inference, supports long context (131k), strong reasoning & tool use, and can run on consumer GPUs (~16GB VRAM). Architecturegpt-oss Parameters20,914,757,184 Context window131,072 tokens Quantization4-bit FormatGGUF Download size11.28 GB Licenseapache-2.0 SHA-256`52f57ab7d3df3ba9173827c1c6832e73375553a846f3e32b49f1ae2daad688d4` `var model = LM.LoadFromModelID("gptoss:20b");` Download gpt-oss-20b-mxfp4.gguf11.3 GB
IBM Granite 3.1 Dense Instructlegacy `granite3.1-dense:2b` superseded by `granite4-h:3b`	text generation, chat, code completion	2.53 B	128K	1.4 GB	apache-2.0
Legacy model. Superseded by `granite4-h:3b`; it remains available for compatibility, but new projects should use the replacement. A long-context instruct model finetuned with a mix of open source and synthetic datasets. Designed for dialogue and text generation tasks. Architecturegranite Parameters2,533,531,648 Context window131,072 tokens Quantization4-bit FormatGGUF Download size1.44 GB Licenseapache-2.0 SHA-256`ba05b36d0a8cebf8ccd13bbbb904bebe182f4854fbcff19cd1ee54bc82bbd298` `var model = LM.LoadFromModelID("granite3.1-dense:2b");` Download granite-3.1-2.5B-Q4_K_M.gguf1.4 GB
IBM Granite 3.1 Dense Instructlegacy `granite3.1-dense:8b` superseded by `granite4-h:7b`	text generation, chat, code completion	8.17 B	128K	4.6 GB	apache-2.0
Legacy model. Superseded by `granite4-h:7b`; it remains available for compatibility, but new projects should use the replacement. An extended-context model optimized for dialogue and code completion tasks. Developed with diverse training data to enhance long-context understanding. Architecturegranite Parameters8,170,848,256 Context window131,072 tokens Quantization4-bit FormatGGUF Download size4.6 GB Licenseapache-2.0 SHA-256`d1ada98d7b274fc6b119bd19b8d3536cd006544e9aae06db6f8b2c256d584044` `var model = LM.LoadFromModelID("granite3.1-dense:8b");` Download granite-3.1-8.2B-Q4_K_M.gguf4.6 GB
IBM Granite 3.3 Instructlegacy `granite3.3:2b` superseded by `granite4-h:3b`	text generation, chat, code completion	2.53 B	128K	1.4 GB	apache-2.0
Legacy model. Superseded by `granite4-h:3b`; it remains available for compatibility, but new projects should use the replacement. A long-context instruct model finetuned with a mix of open source and synthetic datasets. Designed for dialogue and text generation tasks. Architecturegranite Parameters2,533,539,840 Context window131,072 tokens Quantization4-bit FormatGGUF Download size1.44 GB Licenseapache-2.0 SHA-256`dbe4dd51bd6c1e39f96c831bf086454c9b313bd1c279ebb7166f2a37d86598da` `var model = LM.LoadFromModelID("granite3.3:2b");` Download granite-3.3-8B-Instruct-Q4_K_M.gguf1.4 GB
IBM Granite 3.3 Instructlegacy `granite3.3:8b` superseded by `granite4-h:7b`	text generation, chat, code completion	8.17 B	128K	4.6 GB	apache-2.0
Legacy model. Superseded by `granite4-h:7b`; it remains available for compatibility, but new projects should use the replacement. An extended-context model optimized for dialogue and code completion tasks. Developed with diverse training data to enhance long-context understanding. Architecturegranite Parameters8,170,864,640 Context window131,072 tokens Quantization4-bit FormatGGUF Download size4.6 GB Licenseapache-2.0 SHA-256`1c890e740d7ecb010716a858eda315c01ac5bb0edfaf68bf17118868a26bb8ff` `var model = LM.LoadFromModelID("granite3.3:8b");` Download granite-3.3-2B-Instruct-Q4_K_M.gguf4.6 GB
IBM Granite 4 Micro Instruct `granite4-h:3b`	text generation, chat, code completion, tool calling	3.19 B	1024K	1.8 GB	apache-2.0
Hybrid long-context instruct model (Mamba-2 + attention) finetuned from Granite-4.0-H-Micro-Base with SFT, RL alignment, and model merging. Delivers stronger instruction following and robust tool/function calling in multilingual dialog (en, de, es, fr, ja, pt, ar, cs, it, ko, nl, zh), with 1M-token context for enterprise assistants. Excels at summarization, classification, extraction, QA/RAG, and code, including FIM, and supports structured chat templates and OpenAI-style tool schemas. Architecturegranitehybrid Parameters3,191,396,096 Context window1,048,576 tokens Quantization4-bit FormatGGUF Download size1.81 GB Licenseapache-2.0 SHA-256`dbe7b747aa49340f80629811652636b55f4ca4cbbb92ee7e17c442d8a1130566` `var model = LM.LoadFromModelID("granite4-h:3b");` Download Granite-4.0-H-Micro-3.2B-Q4_K_M.gguf1.8 GB
IBM Granite 4 Tiny Instruct `granite4-h:7b`	text generation, chat, code completion, tool calling	6.94 B	1024K	3.9 GB	apache-2.0
Granite-4.0-H-Tiny is an ~7B-parameter hybrid (attention + Mamba2) MoE decoder with a 1M-token context, instruction-tuned (SFT, RL alignment, model merging) for enterprise assistants. It improves instruction following and tool/function calling, supports multilingual dialog (en, de, es, fr, ja, pt, ar, cs, it, ko, nl, zh), and excels at summarization, classification, extraction, QA/RAG, and code/FIM tasks. Architecturegranitehybrid Parameters6,939,037,248 Context window1,048,576 tokens Quantization4-bit FormatGGUF Download size3.94 GB Licenseapache-2.0 SHA-256`75234a50a38235dd4c891dae1a702ccf47a5d89da751d38f90a43be4794f18fb` `var model = LM.LoadFromModelID("granite4-h:7b");` Download Granite-4.0-H-Tiny-64x994M-Q4_K_M.gguf3.9 GB
Microsoft Harrier OSS v1 `harrier-oss:0.6b` MTEB 69.0	embeddings	596 M	32K	340 MB	mit
Microsoft Harrier OSS v1 is a compact multilingual embedding model producing 1024-dimensional instruction-aware embeddings with last-token pooling, covering 90+ languages for semantic search, retrieval, similarity, and bitext mining with long-context support. Architectureqwen3 Parameters596,049,920 Context window32,768 tokens Quantization4-bit FormatGGUF Download size340.08 MB Licensemit MTEB score69.0 · leaderboard SHA-256`ec05d8ceac1302647b1227ab318bd5787e3d58fe6433bc3a4ef7d0730b5e10cd` `var model = LM.LoadFromModelID("harrier-oss:0.6b");` Download harrier-oss-v1-0.6B-Q4_K_M.gguf340 MB
LightOn LightOnOCR 2 `lightonocr-2:1b`	text generation, chat, ocr	1.01 B	16K	628 MB	apache-2.0
LightOnOCR-2-1B is an efficient end-to-end 1B vision-language model for converting documents (PDFs, scans, images) into clean, naturally ordered text, refined with RLVR training for maximum accuracy across tables, receipts, forms, multi-column layouts, and math notation. Architectureqwen3 Parameters1,005,647,872 Context window16,384 tokens Quantization4-bit FormatGGUF Download size628.05 MB Licenseapache-2.0 SHA-256`a3ba5bca4cf4c37b7a269dbe5ee4cb495e6ee3bb4cd1970474ed233b8488240b` `var model = LM.LoadFromModelID("lightonocr-2:1b");` Download lightonocr-2-1b-Q4_K_M.lmk628 MB
LightOn LightOnOCR 2 BBox Soup `lightonocr-2-bbox:1b`	text generation, chat, vision, ocr	1.01 B	16K	628 MB	apache-2.0
LightOnOCR-2-1B-bbox-soup is a merged 1B vision-language model combining OCR and bounding box detection capabilities. Converts documents (PDFs, scans, images) into clean text while localizing embedded images with bounding box coordinates, supporting tables, receipts, forms, multi-column layouts, math notation, and 11+ languages. Architectureqwen3 Parameters1,005,647,872 Context window16,384 tokens Quantization4-bit FormatLMK Download size628.05 MB Licenseapache-2.0 SHA-256`04af6b5928e9b861d10c369e87b541935eccfcf707e9b1e2fbb3f28c82a810ac` `var model = LM.LoadFromModelID("lightonocr-2-bbox:1b");` Download LightOnOCR-2-1B-bbox-soup-Q4_K_M.lmk628 MB
LightOn LightOnOCR 1025legacy `lightonocr1025:1b` superseded by `lightonocr-2:1b`	text generation, chat, ocr	1.16 B	8K	710 MB	apache-2.0
Legacy model. Superseded by `lightonocr-2:1b`; it remains available for compatibility, but new projects should use the replacement. LightOnOCR-1B-1025 is a compact, end-to-end vision–language model for high-accuracy OCR and document understanding, delivering fast, layout-aware text extraction from complex documents. Architectureqwen3 Parameters1,161,230,336 Context window8,192 tokens Quantization4-bit FormatLMK Download size710.36 MB Licenseapache-2.0 SHA-256`42646c9408fd59f342dec352bfc3c84c2255b03e79500506ee9d320e2f66be37` `var model = LM.LoadFromModelID("lightonocr1025:1b");` Download lightonocr-1b-1025-Q4_K_M.lmk710 MB
Meta Llama 3.1 Instruct `llama3.1`	text generation, chat, tool calling	8.03 B	128K	4.6 GB	llama3.1
A multilingual generative model optimized for dialogue and text generation tasks. Designed for robust performance on common benchmarks. Architecturellama Parameters8,030,261,312 Context window131,072 tokens Quantization4-bit FormatGGUF Download size4.58 GB Licensellama3.1 SHA-256`ad00fe50a62d1e009b4e06cd57ab55c9a30cbf5e7f183de09115d75ada73bd5b` `var model = LM.LoadFromModelID("llama3.1");` Download Llama-3.1-8B-Instruct-Q4_K_M.gguf4.6 GB
Meta Llama 3.2 Instruct `llama3.2:1b`	text generation, chat	1.24 B	128K	770 MB	llama3.2
A multilingual instruct-tuned model optimized for dialogue, retrieval, and summarization tasks. Architecturellama Parameters1,235,814,432 Context window131,072 tokens Quantization4-bit FormatGGUF Download size770.28 MB Licensellama3.2 SHA-256`88725e821cf35f1a0dbeaa4a3bebeb91e6c6b6a9d50f808ab42d64233284cce1` `var model = LM.LoadFromModelID("llama3.2:1b");` Download Llama-3.2-1B-Instruct-Q4_K_M.gguf770 MB
Meta Llama 3.2 Instruct `llama3.2:3b`	text generation, chat	3.21 B	128K	1.9 GB	llama3.2
A multilingual dialogue model with robust text generation and summarization capabilities. Architecturellama Parameters3,212,749,888 Context window131,072 tokens Quantization4-bit FormatGGUF Download size1.88 GB Licensellama3.2 SHA-256`6810bf3cce69d440a22b85a3b3e28f57c868f1c98686abd995f1dc5d9b955cfe` `var model = LM.LoadFromModelID("llama3.2:3b");` Download Llama-3.2-3B-Instruct-Q4_K_M.gguf1.9 GB
Meta Llama 3.3 Instruct `llama3.3`	text generation, chat, code completion, math, tool calling	70.6 B	128K	39.6 GB	llama3.3
A large multilingual generative model optimized for dialogue, text tasks, code completion, and mathematical reasoning with extended context support. Architecturellama Parameters70,553,706,560 Context window131,072 tokens Quantization4-bit FormatGGUF Download size39.6 GB Licensellama3.3 SHA-256`57f78fe3b141afa56406278265656524c51c9837edb3537ad43708b6d4ecc04d` `var model = LM.LoadFromModelID("llama3.3");` Download Llama-3.3-70B-Instruct-Q4_K_M.gguf39.6 GB
LM-Kit Sarcasm Detection V1 `lmkit-sarcasm-detection`	sentiment	1.1 B	2K	637 MB	lm-kit
Optimized for detecting sarcasm in English text within the LM-Kit framework. Suitable for CPU-based inference. Architecturellama Parameters1,100,048,384 Context window2,048 tokens Quantization4-bit FormatGGUF Download size636.88 MB Licenselm-kit SHA-256`cc82abd224dba9c689b19d368db6078d6167ca84897b21870d7d6a2c0f09d7d0` `var model = LM.LoadFromModelID("lmkit-sarcasm-detection");` Download LM-Kit.Sarcasm_Detection-TinyLlama-1.1B-1T-OpenOrca-en-q4.gguf637 MB
LM-Kit Sentiment Analysis V2 `lmkit-sentiment-analysis`	sentiment	1.24 B	128K	770 MB	lm-kit
Designed for multilingual sentiment analysis tasks, this LM-Kit model is optimized for efficient CPU-based inference. Architecturellama Parameters1,235,814,432 Context window131,072 tokens Quantization4-bit FormatGGUF Download size770.28 MB Licenselm-kit SHA-256`e12f4abf6453a8431985ce1d6350c265cd58b25210156a917e3608c850fd7add` `var model = LM.LoadFromModelID("lmkit-sentiment-analysis");` Download lm-kit-sentiment-analysis-2.0-1b-q4.gguf770 MB
LM-Kit Tasks Preview `lmkit-tasks:4b-preview`	text generation, chat, code completion, math, vision	3.88 B	128K	3.1 GB	lmkit
A 4B-parameter Gemma3-based model optimized for LM-Kit tasks. Achieves state-of-the-art performance in classification, structured data extraction, language detection, and sentiment analysis, while also supporting chat, embeddings, text generation, code completion, math reasoning, and vision understanding. Designed for seamless integration into LM-Kit pipelines to deliver efficient, reliable, and high-quality results across domains. Architecturegemma3 Parameters3,880,099,328 Context window131,072 tokens Quantization4-bit FormatLMK Download size3.09 GB Licenselmkit SHA-256`3ec9fe4622e2d9a050b3d2c7d2244a911aab75372b04a7bc30bb72a05bdd645c` `var model = LM.LoadFromModelID("lmkit-tasks:4b-preview");` Download lmkit-tasks-4b-preview.lmk3.1 GB
Mistral Magistral Small 1.1legacy `magistral-small` superseded by `magistral-small1.2`	text generation, chat, code completion, math, reasoning	23.6 B	40K	13.3 GB	apache-2.0
Legacy model. Superseded by `magistral-small1.2`; it remains available for compatibility, but new projects should use the replacement. Building upon Mistral Small 3.1 (2503), with added reasoning capabilities, undergoing SFT from Magistral Medium traces and RL on top, it's a small, efficient reasoning model with 24B parameters. Architecturellama Parameters23,572,403,200 Context window40,960 tokens Quantization4-bit FormatGGUF Download size13.35 GB Licenseapache-2.0 SHA-256`7680ba6895d405340f1461cb835a147055689a37d88b193cc5a365aaea76da9e` `var model = LM.LoadFromModelID("magistral-small");` Download Magistral-Small-2506-Q4_K_M.gguf13.3 GB
Mistral Magistral Small 1.2 `magistral-small1.2`	text generation, chat, math, reasoning, tool calling	23.6 B	40K	13.3 GB	apache-2.0
Magistral Small 1.2 (2509) builds upon Mistral Small 3.2 (2506) with added reasoning via SFT from Magistral Medium traces and RL, special [THINK]/[/THINK] tokens, and a 128K context. This GGUF release is text-only (no vision encoder) and should be paired with mistral-common for the correct chat template. Architecturellama Parameters23,572,403,200 Context window40,960 tokens Quantization4-bit FormatGGUF Download size13.35 GB Licenseapache-2.0 SHA-256`d3a024d29e0e8f35d9353f5d4f08fc3715406835c8ae1328ce2f3bd212a43434` `var model = LM.LoadFromModelID("magistral-small1.2");` Download Magistral-Small-2509-Q4_K_M.gguf13.3 GB
OpenBMB MiniCPM o 2.6 Visionlegacy `minicpm-o` superseded by `minicpm-o-45`	text generation, chat, vision	8.12 B	32K	5.0 GB	OpenBMB
Legacy model. Superseded by `minicpm-o-45`; it remains available for compatibility, but new projects should use the replacement. An end-to-end multimodal model supporting real-time speech, image, and text understanding. Offers enhanced performance for conversational tasks. Architectureqwen2 Parameters8,116,736,752 Context window32,768 tokens Quantization4-bit FormatLMK Download size5 GB LicenseOpenBMB SHA-256`6fd17ed1f46bfcddb5a3482dd882dd022a46aa8c33cb93d75f809cd4d118ab53` `var model = LM.LoadFromModelID("minicpm-o");` Download MiniCPM-o-V-2.6-Q4_K_M.lmk5.0 GB
OpenBMB MiniCPM o 4.5 Vision `minicpm-o-45`	text generation, chat, vision	8.72 B	40K	5.3 GB	apache-2.0
MiniCPM-o 4.5 is an end-to-end multimodal model built on SigLip2, Whisper-medium, CosyVoice2, and Qwen3-8B, delivering GPT-4o-level vision-language performance with 9B parameters. Supports instruct and thinking modes, high-resolution images up to 1.8M pixels, high-FPS video understanding, real-time speech conversation, and full-duplex multimodal live streaming. Architectureqwen3 Parameters8,715,965,680 Context window40,960 tokens Quantization4-bit FormatLMK Download size5.3 GB Licenseapache-2.0 SHA-256`234387b78e1306cecabc51d5afd28fa44077f14fff4a2b383586099749b1d090` `var model = LM.LoadFromModelID("minicpm-o-45");` Download minicpm-o-4.5-Q4_K_M.lmk5.3 GB
OpenBMB MiniCPM 2.6 Visionlegacy `minicpm-v` superseded by `minicpm-o-45`	text generation, chat, vision	8.12 B	32K	5.0 GB	OpenBMB
Legacy model. Superseded by `minicpm-o-45`; it remains available for compatibility, but new projects should use the replacement. A multimodal model designed for vision and text tasks, built upon SigLip and Qwen architectures. Evaluate performance against current benchmarks. Architectureqwen2 Parameters8,116,736,752 Context window32,768 tokens Quantization4-bit FormatLMK Download size5 GB LicenseOpenBMB SHA-256`a10b1aa434899ea0bd5bb5e281f622fed0b02434241d53435fce05773fa7cfa8` `var model = LM.LoadFromModelID("minicpm-v");` Download MiniCPM-V-2.6-Q4_K_M.lmk5.0 GB
OpenBMB MiniCPM-V 4.5 `minicpm-v-45`	text generation, chat, vision, ocr	8.72 B	40K	5.7 GB	OpenBMB
MiniCPM-V 4.5 is a state-of-the-art multimodal LLM built on Qwen3-8B and SigLIP2-400M. It delivers GPT-4o-level performance for single-image, multi-image, and high-FPS video understanding on local devices. The model supports controllable fast/deep thinking, real-time speech and text comprehension, strong OCR and document parsing (up to 1.8M pixels), and multilingual capabilities in 30+ languages. Optimized for efficiency, it enables CPU inference, mobile deployment, and scalable usage through formats like LMK, GGUF, and AWQ. Architectureqwen3 Parameters8,715,965,680 Context window40,960 tokens Quantization4-bit FormatLMK Download size5.7 GB LicenseOpenBMB SHA-256`000c56809f033e53637f364461cfadb8c4aa09e533a3fde66de39cbb41bf5cb7` `var model = LM.LoadFromModelID("minicpm-v-45");` Download minicpm-v-4.5-8b.lmk5.7 GB
Mistral Ministral 3 `ministral3:3b`	text generation, chat, math, vision, tool calling	3.85 B	256K	2.4 GB	apache-2.0
Smallest member of the Ministral 3 family, an edge-optimized multilingual instruct model with a 256K context window and solid reasoning and code capabilities for constrained hardware. Architecturemistral3 Parameters3,849,093,120 Context window262,144 tokens Quantization4-bit FormatLMK Download size2.42 GB Licenseapache-2.0 SHA-256`300ce4373c14a3c1d68e37a7f4537b98776eb8ccac50ff32b30cc8ae6191ee96` `var model = LM.LoadFromModelID("ministral3:3b");` Download ministral-3-3b-instruct-Q4_K_M.lmk2.4 GB
Mistral Ministral 3 `ministral3:8b`	text generation, chat, math, vision, tool calling	8.92 B	256K	5.3 GB	apache-2.0
Mid-sized Ministral 3 variant that balances quality and cost, offering strong multilingual reasoning, math, and code performance with a 256K context while remaining practical for single-GPU and edge deployments. Architecturemistral3 Parameters8,918,030,336 Context window262,144 tokens Quantization4-bit FormatLMK Download size5.27 GB Licenseapache-2.0 SHA-256`4bd03de58774150e6d19e9a1f0e4f1e010784ac7b98801c35144f5fd796c81c8` `var model = LM.LoadFromModelID("ministral3:8b");` Download ministral-3-8b-instruct-Q4_K_M.lmk5.3 GB
Mistral Ministral 3 `ministral3:14b`	text generation, chat, math, vision, tool calling	13.9 B	256K	8.1 GB	apache-2.0
Flagship Ministral 3 instruct model delivering frontier-level multilingual reasoning, math, and code performance in a 256K-token context, with design tuned for efficient edge and single-GPU deployment. Architecturemistral3 Parameters13,945,036,800 Context window262,144 tokens Quantization4-bit FormatLMK Download size8.11 GB Licenseapache-2.0 SHA-256`032759d8b1814347ae571a2dd6d24c2d36760141f07756a6559bb77a17a9e821` `var model = LM.LoadFromModelID("ministral3:14b");` Download ministral-3-14b-instruct-Q4_K_M.lmk8.1 GB
Mistral Nemo Instruct 2407legacy `mistral-nemo` superseded by `ministral3:8b`	text generation, chat	12.2 B	1000K	7.0 GB	apache-2.0
Legacy model. Superseded by `ministral3:8b`; it remains available for compatibility, but new projects should use the replacement. An instruct-tuned variant developed in collaboration with NVIDIA, balancing model size with performance for conversational tasks. Architecturellama Parameters12,247,782,400 Context window1,024,000 tokens Quantization4-bit FormatGGUF Download size6.96 GB Licenseapache-2.0 SHA-256`579ab8f5178f5900d0c4e14534929aa0dba97e3f97be76b31ebe537ffd6cf169` `var model = LM.LoadFromModelID("mistral-nemo");` Download Mistral-Nemo-2407-12.2B-Instruct-Q4_K_M.gguf7.0 GB
Mistral Small Instruct 2501legacy `mistral-small` superseded by `mistral-small3.2`	text generation, chat, code completion, math	23.6 B	32K	13.3 GB	apache-2.0
Legacy model. Superseded by `mistral-small3.2`; it remains available for compatibility, but new projects should use the replacement. Optimized for local deployment, this model balances parameter count and performance for chat and code tasks. Architecturellama Parameters23,572,403,200 Context window32,768 tokens Quantization4-bit FormatGGUF Download size13.35 GB Licenseapache-2.0 SHA-256`4395b5c6136e29e9b11bdba2ee189302ad45dd5c3ef45073b729f077b8f0cec8` `var model = LM.LoadFromModelID("mistral-small");` Download Mistral-Small-Instruct-2501-24B-Q4_K_M.gguf13.3 GB
Mistral Small 3.1 Instruct 2503legacy `mistral-small3.1` superseded by `mistral-small3.2`	text generation, chat, code completion, math	23.6 B	128K	13.3 GB	apache-2.0
Legacy model. Superseded by `mistral-small3.2`; it remains available for compatibility, but new projects should use the replacement. Mistral Small 3.1 (24B) enhances Mistral Small 3 with advanced vision, 128k context, multilingual support, agentic features, and efficient local deployment. Architecturellama Parameters23,572,403,200 Context window131,072 tokens Quantization4-bit FormatGGUF Download size13.35 GB Licenseapache-2.0 SHA-256`68922ff3a311c81bc4e983f86e665a12213ee84710c210522f10e65ce980bda7` `var model = LM.LoadFromModelID("mistral-small3.1");` Download Mistral-Small-3.1-24B-Instruct-2503-Q4_K_M.gguf13.3 GB
Mistral Small 3.2 Instruct 2503 `mistral-small3.2`	text generation, chat, code completion, math, tool calling	23.6 B	128K	13.3 GB	apache-2.0
Mistral Small 3.2 (24B) enhances Mistral Small 3 with advanced vision, 128k context, multilingual support, agentic features, and efficient local deployment. Architecturellama Parameters23,572,403,200 Context window131,072 tokens Quantization4-bit FormatGGUF Download size13.35 GB Licenseapache-2.0 SHA-256`e1e9a516a90387ec98bb9c45c37dbb1478008d1fa46b216cca893cf008d92c29` `var model = LM.LoadFromModelID("mistral-small3.2");` Download Mistral-Small-3.2-24B-Instruct-2506-Q4_K_M.gguf13.3 GB
NVIDIA Nemotron 3 Nano `nemotron3-nano`	text generation, chat, code completion, math, reasoning, tool calling	31.6 B	1024K	22.8 GB	nvidia-open-model-license
A hybrid Mamba-2/Transformer MoE reasoning model (30B total, 3.5B active per token) trained for both thinking and non-thinking modes. Excels at agentic tasks, math, code, tool calling, and instruction following across English, German, Spanish, French, Italian, and Japanese. Architecturenemotron_h_moe Parameters31,577,940,288 Context window1,048,576 tokens Quantization4-bit FormatGGUF Download size22.83 GB Licensenvidia-open-model-license SHA-256`90bedf225d55b3a524db3f6e2ba202c13adb389ddab261a233082e9e715a85e0` `var model = LM.LoadFromModelID("nemotron3-nano");` Download NVIDIA-Nemotron-3-Nano-30B-A3B-Q4_K_M.gguf22.8 GB
NVIDIA Nemotron 3 Nano `nemotron3-nano:4b`	text generation, chat, code completion, math, reasoning, tool calling	3.97 B	256K	2.6 GB	nvidia-nemotron-open-model-license
An edge-ready hybrid Mamba-2/Transformer model (3.97B parameters) compressed from Nemotron Nano 9B using the Nemotron Elastic framework. Designed for both reasoning and non-reasoning modes, excelling at agentic tasks, math, code, tool calling, and instruction following. Intended for edge platforms such as Jetson Thor, GeForce RTX, and DGX Spark. Architecturenemotron_h Parameters3,970,000,000 Context window262,144 tokens Quantization4-bit FormatGGUF Download size2.64 GB Licensenvidia-nemotron-open-model-license SHA-256`be5d9a656a51922f24f1f09a759cebb694e1f5d9728bf0ef9f8c972c5a0b5ef2` `var model = LM.LoadFromModelID("nemotron3-nano:4b");` Download NVIDIA-Nemotron3-Nano-4B-Q4_K_M.gguf2.6 GB
Nomic embed text v1.5 `nomic-embed-text`	embeddings	137 M	2K	86 MB	apache-2.0
Provides flexible production embeddings using Matryoshka Representation Learning. Architecturenomic-bert Parameters136,731,648 Context window2,048 tokens Quantization4-bit FormatGGUF Download size85.86 MB Licenseapache-2.0 SHA-256`1a60949a331b30bb754ad60b7bdff80d8e563a56b3f7f3f1aed68db8c143003e` `var model = LM.LoadFromModelID("nomic-embed-text");` Download nomic-embed-text-1.5-Q4_K_M.gguf86 MB
Nomic embed vision v1.5 `nomic-embed-vision`	image embeddings	92 M	197	92 MB	apache-2.0
ViT-B/16-based image embedding model trained on 1.5B image-text pairs using Matryoshka Representation Learning. Outputs 768-dim embeddings aligned with Nomic Embed Text v1.5 for multimodal search, retrieval, and zero-shot classification. ArchitectureViT-B/16 Parameters92,384,769 Context window197 tokens Quantization8-bit FormatONNX Download size92.26 MB Licenseapache-2.0 SHA-256`4f6f6a765625a4b74ec3e62141b7b83e1db1fb904afeda1fa00c1fefefbcc714` `var model = LM.LoadFromModelID("nomic-embed-vision");` Download nomic-embed-vision-1.5-Q8.lmk92 MB
PaddlePaddle PaddleOCR VL 1.5legacy `paddleocr-vl:0.9b` superseded by `paddleocr-vl-1.6:0.9b`	text generation, chat, vision, ocr	906 M	128K	751 MB	apache-2.0
Legacy model. Superseded by `paddleocr-vl-1.6:0.9b`; it remains available for compatibility, but new projects should use the replacement. PaddleOCR-VL 1.5 is an ultra-compact 0.9B vision-language model built on ERNIE-4.5-0.3B for high-accuracy document understanding. It supports OCR, table recognition, formula recognition, chart recognition, text spotting, and seal recognition, achieving 94.5% accuracy on OmniDocBench v1.5 across challenging real-world scenarios. Architecturepaddleocr Parameters905,601,648 Context window131,072 tokens Quantization4-bit FormatLMK Download size750.86 MB Licenseapache-2.0 SHA-256`4dbcb6c3ae742dcc1e973ca5234071cc816b29987a0e5f6a39ffe1cc79273c0a` `var model = LM.LoadFromModelID("paddleocr-vl:0.9b");` Download paddle-ocr-vl-1.5-Q4_K_M.lmk751 MB
PaddlePaddle PaddleOCR VL 1.6 `paddleocr-vl-1.6:0.9b`	text generation, chat, vision, ocr	906 M	128K	725 MB	apache-2.0
PaddleOCR-VL 1.6 is an ultra-compact 0.9B vision-language model built on ERNIE-4.5-0.3B for high-accuracy document parsing. Architecturally compatible with version 1.5 for drop-in migration, it supports OCR, table recognition, formula recognition, chart recognition, text spotting, seal recognition, layout analysis, and ancient document and rare character recognition. It reaches 96.33% on OmniDocBench v1.6 (state of the art) and sets new records on Real5-OmniDocBench across scanning, warping, screen-photography, illumination, and skew scenarios, with notable gains in table, chart, seal, and Chinese ancient document recognition over version 1.5. Architecturepaddleocr Parameters905,601,648 Context window131,072 tokens Quantization4-bit FormatLMK Download size724.82 MB Licenseapache-2.0 SHA-256`3873b666944a4327c882cbb95d4539e0874c79b13f0034e2868734f1c076c520` `var model = LM.LoadFromModelID("paddleocr-vl-1.6:0.9b");` Download paddle-ocr-vl-1.6-Q4_K_M.lmk725 MB
Microsoft Phi 3.5 Mini Instructlegacy `phi3.5` superseded by `phi4-mini`	text generation, chat	3.82 B	128K	2.2 GB	mit
Legacy model. Superseded by `phi4-mini`; it remains available for compatibility, but new projects should use the replacement. A lightweight model optimized for reasoning-dense tasks and extended context support. Designed for efficient instruction following. Architecturephi3 Parameters3,821,079,648 Context window131,072 tokens Quantization4-bit FormatGGUF Download size2.23 GB Licensemit SHA-256`782c34ae79564d1d92bd44dec233182559b3ecf6fedee44417e2a28c89bd0721` `var model = LM.LoadFromModelID("phi3.5");` Download Phi-3.5-mini-Instruct-Q4_K_M.gguf2.2 GB
Microsoft Phi 4 Instruct `phi4`	text generation, chat, math, tool calling	14.7 B	16K	8.4 GB	mit
An enhanced generative model trained on a diverse dataset to improve instruction adherence and reasoning capabilities. Architecturephi3 Parameters14,659,507,200 Context window16,384 tokens Quantization4-bit FormatGGUF Download size8.43 GB Licensemit SHA-256`03af8f5c5a87d526047f5c20c99e32bbafd5db6dbfdee8d498d0fe1a3c45af55` `var model = LM.LoadFromModelID("phi4");` Download Phi-4-14.7B-Instruct-Q4_K_M.gguf8.4 GB
Microsoft Phi 4 Mini Instruct `phi4-mini`	text generation, chat, tool calling	3.84 B	128K	2.3 GB	mit
A lightweight open model from the Phi-4 family that uses synthetic and curated public data for reasoning-dense outputs, supports a 128K token context, and is enhanced through fine-tuning and preference optimization for precise instruction adherence and robust safety. Architecturephi3 Parameters3,836,021,856 Context window131,072 tokens Quantization4-bit FormatGGUF Download size2.32 GB Licensemit SHA-256`556492e72efc8d33406b236830ad38d25669482ea7ad91fc643de237e942b9f9` `var model = LM.LoadFromModelID("phi4-mini");` Download Phi-4-mini-Instruct-Q4_K_M.gguf2.3 GB
Mistral Pixtral `pixtral`	text generation, chat, vision	12.7 B	1000K	7.4 GB	apache-2.0
Pixtral 12B is a natively multimodal model combining a 12 B parameter decoder with a 400 M vision encoder, trained on interleaved image–text data for variable image sizes, offering state-of-the-art performance in its weight class across multimodal and text-only benchmarks and supporting ultra-long 128 k sequence lengths. Architecturellama Parameters12,682,744,832 Context window1,024,000 tokens Quantization4-bit FormatLMK Download size7.39 GB Licenseapache-2.0 SHA-256`28d42e60b5f33765ac6f3882abc4c7fd9f5a7955910ff117c13dbfc5aa6bf159` `var model = LM.LoadFromModelID("pixtral");` Download pixtral-12B-Q4_K_M.lmk7.4 GB
Alibaba Qwen 2 Vision Instructlegacy `qwen2-vl:2b` superseded by `qwen3.5:2b`	text generation, chat, vision	2.21 B	32K	1.3 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:2b`; it remains available for compatibility, but new projects should use the replacement. A multilingual vision-language model featuring dynamic resolution processing for advanced image and long-video understanding. Architectureqwen2vl Parameters2,208,985,700 Context window32,768 tokens Quantization4-bit FormatLMK Download size1.27 GB Licenseapache-2.0 SHA-256`b4e546acfd2271f5a0960b64445cae1091e5fc4192d74db72ae57c28729bd0b8` `var model = LM.LoadFromModelID("qwen2-vl:2b");` Download Qwen2-VL-2B-Instruct-Q4_K_M.lmk1.3 GB
Alibaba Qwen 2 Vision Instructlegacy `qwen2-vl:8b` superseded by `qwen3.5:9b`	text generation, chat, vision	8.29 B	32K	4.7 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:9b`; it remains available for compatibility, but new projects should use the replacement. An extended variant in the Qwen 2 Vision family for multilingual vision-language tasks, including advanced video analysis. Architectureqwen2vl Parameters8,291,375,716 Context window32,768 tokens Quantization4-bit FormatLMK Download size4.72 GB Licenseapache-2.0 SHA-256`90b3eb60611559ba7521590ecccdf1d2a4dfab007566221c6a42f19b91b48686` `var model = LM.LoadFromModelID("qwen2-vl:8b");` Download Qwen2-VL-8.3B-Instruct-Q4_K_M.lmk4.7 GB
Alibaba Qwen 2.5 Instructlegacy `qwen2.5:0.5b` superseded by `qwen3.5:0.8b`	text generation, chat	494 M	32K	379 MB	apache-2.0
Legacy model. Superseded by `qwen3.5:0.8b`; it remains available for compatibility, but new projects should use the replacement. A compact variant from the Alibaba Qwen 2.5 family, optimized for instruction following across chat, embeddings, and text generation tasks. Architectureqwen2 Parameters494,032,768 Context window32,768 tokens Quantization4-bit FormatGGUF Download size379.38 MB Licenseapache-2.0 SHA-256`09b44ff0ef0a160ffe50778c0828754201bb3a40522a941839c23acfbc9ceec0` `var model = LM.LoadFromModelID("qwen2.5:0.5b");` Download Qwen-2.5-0.5B-Instruct-Q4_K_M.gguf379 MB
Alibaba Qwen 2.5 Instructlegacy `qwen2.5:3b` superseded by `qwen3.5:4b`	text generation, chat	3.09 B	32K	1.8 GB	qwen-research
Legacy model. Superseded by `qwen3.5:4b`; it remains available for compatibility, but new projects should use the replacement. A mid-sized model from the Alibaba Qwen 2.5 series, designed for diverse tasks including chat, embeddings, and text generation. Performance should be evaluated relative to current benchmarks. Architectureqwen2 Parameters3,085,938,688 Context window32,768 tokens Quantization4-bit FormatGGUF Download size1.8 GB Licenseqwen-research SHA-256`fb88cca2303e7f7d4d52679d633efe66d9c3e3555573b4444abe5ab8af4a97f7` `var model = LM.LoadFromModelID("qwen2.5:3b");` Download Qwen-2.5-3.1B-Instruct-Q4_K_M.gguf1.8 GB
Alibaba Qwen 2.5 Instructlegacy `qwen2.5:7b` superseded by `qwen3.5:9b`	text generation, chat	7.62 B	32K	4.4 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:9b`; it remains available for compatibility, but new projects should use the replacement. A larger variant from the Alibaba Qwen 2.5 series that supports extended context and multiple tasks including chat, embeddings, and text generation. Architectureqwen2 Parameters7,615,616,512 Context window32,768 tokens Quantization4-bit FormatGGUF Download size4.36 GB Licenseapache-2.0 SHA-256`2bf11b8a7d566bddfcc2b222ed7b918afc51239c5f919532de8b9403981ad866` `var model = LM.LoadFromModelID("qwen2.5:7b");` Download Qwen-2.5-7B-Instruct-Q4_K_M.gguf4.4 GB
Alibaba Qwen 2.5 Vision Instructlegacy `qwen2.5-vl:3b` superseded by `qwen3.5:2b`	text generation, chat, vision	3.75 B	125K	2.6 GB	qwen research license
Legacy model. Superseded by `qwen3.5:2b`; it remains available for compatibility, but new projects should use the replacement. Qwen2.5 VL 3B Instruct is a compact vision-language chat model that delivers advanced object and text/chart understanding, agentic tool-driven interactions, long-video event localization, precise visual grounding with JSON outputs, and structured data extraction, powered by an optimized ViT encoder with dynamic temporal training. Architectureqwen2vl Parameters3,754,622,976 Context window128,000 tokens Quantization4-bit FormatLMK Download size2.58 GB Licenseqwen research license SHA-256`78fee4fde9f7fd93e1365cae46668184a259b1bd2a3169915a4a1e7495f859f8` `var model = LM.LoadFromModelID("qwen2.5-vl:3b");` Download Qwen2.5-VL-3B-Instruct-Q4_K_M.lmk2.6 GB
Alibaba Qwen 2.5 Vision Instructlegacy `qwen2.5-vl:7b` superseded by `qwen3.5:9b`	text generation, chat, vision	8.29 B	125K	5.2 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:9b`; it remains available for compatibility, but new projects should use the replacement. Qwen2.5 VL 7B Instruct is a next-generation vision-language chat model that combines advanced object and text/chart understanding, agentic tool use, long-video event localization, precise visual grounding with JSON outputs, and structured data extraction, all powered by a streamlined ViT encoder with dynamic temporal training. Architectureqwen2vl Parameters8,292,166,656 Context window128,000 tokens Quantization4-bit FormatLMK Download size5.16 GB Licenseapache-2.0 SHA-256`e9a99c7bb06c23bd60594cebf8a881af13f502742df3047eaa3b466c747f7453` `var model = LM.LoadFromModelID("qwen2.5-vl:7b");` Download Qwen2.5-VL-7B-Instruct-Q4_K_M.lmk5.2 GB
Alibaba Qwen 2.5 Vision Instructlegacy `qwen2.5-vl:32b` superseded by `qwen3.6:35b-a3b`	text generation, chat, vision	33.5 B	125K	18.9 GB	apache-2.0
Legacy model. Superseded by `qwen3.6:35b-a3b`; it remains available for compatibility, but new projects should use the replacement. Qwen2.5 VL 32B Instruct is a next-generation vision-language chat model that combines advanced object and text/chart understanding, agentic tool use, long-video event localization, precise visual grounding with JSON outputs, and structured data extraction, all powered by a streamlined ViT encoder with dynamic temporal training. Architectureqwen2vl Parameters33,452,718,336 Context window128,000 tokens Quantization4-bit FormatLMK Download size18.92 GB Licenseapache-2.0 SHA-256`9081e05fc2832177162b9ea8ccde1e0fdb1d8ed429a838527af36de966e2fb92` `var model = LM.LoadFromModelID("qwen2.5-vl:32b");` Download Qwen2.5-VL-32B-Instruct-Q4_K_M.lmk18.9 GB
Alibaba Qwen 3 Instructlegacy `qwen3:0.6b` superseded by `qwen3.5:0.8b`	text generation, chat, reasoning	752 M	40K	462 MB	apache-2.0
Legacy model. Superseded by `qwen3.5:0.8b`; it remains available for compatibility, but new projects should use the replacement. Qwen3 is the latest generation of Qwen large language models, combining dense and MoE architectures with seamless “thinking” vs. “non‐thinking” mode switching to deliver state-of-the-art reasoning, coding, agent integration, and instruction-following across 100+ languages. Architectureqwen3 Parameters751,632,384 Context window40,960 tokens Quantization4-bit FormatGGUF Download size461.79 MB Licenseapache-2.0 SHA-256`2b1a7ed56061ad1275847412f61e8e009ada37ef865dccc25747dcc76eea9811` `var model = LM.LoadFromModelID("qwen3:0.6b");` Download Qwen3-0.6B-Q4_K_M.gguf462 MB
Alibaba Qwen 3 Instructlegacy `qwen3:1.7b` superseded by `qwen3.5:2b`	text generation, chat, reasoning	2.03 B	40K	1.2 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:2b`; it remains available for compatibility, but new projects should use the replacement. Qwen3 is the latest generation of Qwen large language models, combining dense and MoE architectures with seamless “thinking” vs. “non‐thinking” mode switching to deliver state-of-the-art reasoning, coding, agent integration, and instruction-following across 100+ languages. Architectureqwen3 Parameters2,031,739,904 Context window40,960 tokens Quantization4-bit FormatGGUF Download size1.19 GB Licenseapache-2.0 SHA-256`b047d6617eba56dcfa3357566b06807f54b15816faf6182aabd12d7e2378e537` `var model = LM.LoadFromModelID("qwen3:1.7b");` Download Qwen3-1.7B-Q4_K_M.gguf1.2 GB
Alibaba Qwen 3 Instructlegacy `qwen3:4b` superseded by `qwen3.5:4b`	text generation, chat, math, reasoning, tool calling	4.02 B	40K	2.3 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:4b`; it remains available for compatibility, but new projects should use the replacement. Qwen3 is the latest generation of Qwen large language models, combining dense and MoE architectures with seamless “thinking” vs. “non‐thinking” mode switching to deliver state-of-the-art reasoning, coding, agent integration, and instruction-following across 100+ languages. Architectureqwen3 Parameters4,022,468,096 Context window40,960 tokens Quantization4-bit FormatGGUF Download size2.33 GB Licenseapache-2.0 SHA-256`9dbc1e801f001ea316a627bb867fdd192fc3b36046fd69e160155ddc12129dbe` `var model = LM.LoadFromModelID("qwen3:4b");` Download Qwen3-4B-Q4_K_M.gguf2.3 GB
Alibaba Qwen 3 Instructlegacy `qwen3:8b` superseded by `qwen3.5:9b`	text generation, chat, math, reasoning, tool calling	8.19 B	40K	4.7 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:9b`; it remains available for compatibility, but new projects should use the replacement. Qwen3 is the latest generation of Qwen large language models, combining dense and MoE architectures with seamless “thinking” vs. “non‐thinking” mode switching to deliver state-of-the-art reasoning, coding, agent integration, and instruction-following across 100+ languages. Architectureqwen3 Parameters8,190,735,360 Context window40,960 tokens Quantization4-bit FormatGGUF Download size4.68 GB Licenseapache-2.0 SHA-256`b9059e3978453f50a8e9e45a825243abdb8739b2f4623e541fd5a392d9672c0f` `var model = LM.LoadFromModelID("qwen3:8b");` Download Qwen3-8B-Q4_K_M.gguf4.7 GB
Alibaba Qwen 3 Instructlegacy `qwen3:14b` superseded by `qwen3.5:9b`	text generation, chat, math, reasoning, tool calling	14.8 B	40K	8.4 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:9b`; it remains available for compatibility, but new projects should use the replacement. Qwen3 is the latest generation of Qwen large language models, combining dense and MoE architectures with seamless “thinking” vs. “non‐thinking” mode switching to deliver state-of-the-art reasoning, coding, agent integration, and instruction-following across 100+ languages. Architectureqwen3 Parameters14,768,307,200 Context window40,960 tokens Quantization4-bit FormatGGUF Download size8.38 GB Licenseapache-2.0 SHA-256`520369028ee99e4a3ca413a35126337038a8da561927f81322c1b34aed10e03d` `var model = LM.LoadFromModelID("qwen3:14b");` Download Qwen3-14B-Q4_K_M.gguf8.4 GB
Alibaba Qwen 3 Coder Instruct 30B-A3B `qwen3-coder:30b-a3b`	text generation, chat, code completion, tool calling	30.5 B	256K	17.3 GB	apache-2.0
Qwen3-Coder 30B-A3B is a specialized code-focused Mixture-of-Experts model (30.5B total, 3.3B active) with 128 experts, delivering state-of-the-art agentic coding, tool calling, and long-context code understanding with a native 262K context window. Architectureqwen3moe Parameters30,532,122,624 Context window262,144 tokens Quantization4-bit FormatGGUF Download size17.28 GB Licenseapache-2.0 SHA-256`956682fa9d36d4d0e5a80eb90ff8a001f2c48f988a497e565ae4d0c42af4fe44` `var model = LM.LoadFromModelID("qwen3-coder:30b-a3b");` Download Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf17.3 GB
Alibaba Qwen 3 Embedding `qwen3-embedding:0.6b` MTEB 64.33	embeddings	596 M	32K	610 MB	apache-2.0
Lightweight member of the Qwen3 Embedding series, optimized for fast, low-resource semantic search and ranking while preserving strong multilingual and long-context understanding. Architectureqwen3 Parameters595,776,512 Context window32,768 tokens Quantization4-bit FormatGGUF Download size609.54 MB Licenseapache-2.0 MTEB score64.33 · leaderboard SHA-256`b624c62027986bc4181eadcad0cee479916c498d1039f7063195fd4c14803023` `var model = LM.LoadFromModelID("qwen3-embedding:0.6b");` Download Qwen3-Embedding-0.6B-Q4_K_M.gguf610 MB
Alibaba Qwen 3 Embedding `qwen3-embedding:4b` MTEB 69.45	embeddings	4.02 B	40K	2.3 GB	apache-2.0
Mid-size Qwen3 Embedding model offering a strong accuracy–efficiency trade-off for multilingual retrieval, reranking, classification, clustering, and bitext mining with instruction-aware embeddings. Architectureqwen3 Parameters4,021,774,336 Context window40,960 tokens Quantization4-bit FormatGGUF Download size2.33 GB Licenseapache-2.0 MTEB score69.45 · leaderboard SHA-256`ac48f080498db5874835a0c6db52aa8a726f8f88fca1dbbb26fc51d5311acb85` `var model = LM.LoadFromModelID("qwen3-embedding:4b");` Download Qwen3-Embedding-4B-Q4_K_M.gguf2.3 GB
Alibaba Qwen 3 Embedding `qwen3-embedding:8b` MTEB 70.58	embeddings	7.57 B	40K	4.4 GB	apache-2.0
Flagship Qwen3 Embedding model delivering state-of-the-art multilingual and cross-lingual embeddings for dense retrieval, reranking, text/code search, clustering, and classification, with flexible output dimensions. Architectureqwen3 Parameters7,567,295,488 Context window40,960 tokens Quantization4-bit FormatGGUF Download size4.36 GB Licenseapache-2.0 MTEB score70.58 · leaderboard SHA-256`3822cc88f2f3e9a08c4b9bae87261a7e94c503fe0372ad9b6c5b80161886291a` `var model = LM.LoadFromModelID("qwen3-embedding:8b");` Download Qwen3-Embedding-8B-Q4_K_M.gguf4.4 GB
Alibaba Qwen 3 Vision Instructlegacy `qwen3-vl:2b` superseded by `qwen3.5:2b`	text generation, chat, code completion, math, vision, tool calling, ocr	2.13 B	256K	1.3 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:2b`; it remains available for compatibility, but new projects should use the replacement. Qwen3-VL is the latest Qwen vision-language family with stronger text understanding/generation, deeper visual reasoning, native 256K context (expandable), upgraded OCR (32 langs), long-video comprehension, and agentic GUI/tool use. The 2B Instruct edition targets edge devices for multimodal chat, grounding, and document/image understanding. Architectureqwen3vl Parameters2,127,532,032 Context window262,144 tokens Quantization4-bit FormatLMK Download size1.29 GB Licenseapache-2.0 SHA-256`296c414827f80a0205371f2d407169f71457c4b9f9ed49edd1b28e8b9f697ace` `var model = LM.LoadFromModelID("qwen3-vl:2b");` Download qwen3-vl-2b-instruct-Q4_K_M.lmk1.3 GB
Alibaba Qwen 3 Vision Instructlegacy `qwen3-vl:4b` superseded by `qwen3.5:4b`	text generation, chat, code completion, math, vision, tool calling, ocr	2.13 B	256K	2.6 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:4b`; it remains available for compatibility, but new projects should use the replacement. Qwen3-VL 4B Instruct balances efficiency and quality with advanced spatial perception (2D/3D grounding), timestamp-aligned video reasoning, and “visual coding” (HTML/CSS/JS from images). Suited for on-device or small-GPU multimodal assistants, retrieval, and structured understanding. Architectureqwen3vl Parameters2,127,532,032 Context window262,144 tokens Quantization4-bit FormatLMK Download size2.59 GB Licenseapache-2.0 SHA-256`30c7ff027b6b148950533fbc6cf473abc288245af61cd23ca3ae135b6db9f3e8` `var model = LM.LoadFromModelID("qwen3-vl:4b");` Download qwen3-vl-4b-instruct-Q4_K_M.lmk2.6 GB
Alibaba Qwen 3 Vision Instructlegacy `qwen3-vl:8b` superseded by `qwen3.5:9b`	text generation, chat, code completion, math, vision, tool calling, ocr	8.77 B	256K	5.2 GB	apache-2.0
Legacy model. Superseded by `qwen3.5:9b`; it remains available for compatibility, but new projects should use the replacement. Qwen3-VL 8B Instruct is the highest-quality dense variant, delivering state-of-the-art multimodal reasoning, long-horizon video understanding, stronger recognition (celebrities, products, flora/fauna, etc.), and robust agent/tool interaction, ideal for high-fidelity VLM chat and STEM tasks. Architectureqwen3vl Parameters8,767,123,696 Context window262,144 tokens Quantization4-bit FormatLMK Download size5.21 GB Licenseapache-2.0 SHA-256`76031476ae3f8720e01a941e655981c7b6779b5709a746879d771534a3d0ccdf` `var model = LM.LoadFromModelID("qwen3-vl:8b");` Download qwen3-vl-8b-instruct-Q4_K_M.lmk5.2 GB
Alibaba Qwen 3 Vision Instructlegacy `qwen3-vl:30b` superseded by `qwen3.6:35b-a3b`	text generation, chat, code completion, math, vision, tool calling, ocr	31.1 B	256K	17.8 GB	apache-2.0
Legacy model. Superseded by `qwen3.6:35b-a3b`; it remains available for compatibility, but new projects should use the replacement. Qwen3-VL 30B Instruct is the flagship MoE vision-language model in the Qwen3 family, delivering state-of-the-art multimodal reasoning, long-horizon video and document understanding with a native 256K context, precise spatial grounding and timestamp-aligned event localization, expanded OCR in 32 languages, and powerful visual-agent capabilities for GUI control, visual coding, and tool-augmented STEM and math workflows. Architectureqwen3vlmoe Parameters31,070,754,032 Context window262,144 tokens Quantization4-bit FormatLMK Download size17.79 GB Licenseapache-2.0 SHA-256`955041de8ecc63327b6a5fa408ccf03e80e62e663b88f5314aa2901c332b7478` `var model = LM.LoadFromModelID("qwen3-vl:30b");` Download qwen3-vl-30b-instruct-Q4_K_M.lmk17.8 GB
Alibaba Qwen 3.5 `qwen3.5:0.8b`	text generation, chat, vision	853 M	256K	577 MB	apache-2.0
Qwen3.5 0.8B is an ultra-compact dense hybrid model using Gated Delta Networks with gated attention, providing efficient multilingual chat and vision capabilities across 200+ languages with a native 262K context window. Architectureqwen35 Parameters852,985,920 Context window262,144 tokens Quantization4-bit FormatLMK Download size576.61 MB Licenseapache-2.0 SHA-256`a84b4b20ba32565dc91f59f78874f4395dab11f1d2e0115698b4bdb510a966e3` `var model = LM.LoadFromModelID("qwen3.5:0.8b");` Download Qwen3.5-0.8B-Q4_K_M.lmk577 MB
Alibaba Qwen 3.5 `qwen3.5:2b`	text generation, chat, math, vision	2.21 B	256K	1.4 GB	apache-2.0
Qwen3.5 2B is a compact dense hybrid model using Gated Delta Networks with gated attention, providing strong multilingual chat, vision, and math capabilities across 200+ languages with a native 262K context window. Architectureqwen35 Parameters2,213,241,664 Context window262,144 tokens Quantization4-bit FormatLMK Download size1.4 GB Licenseapache-2.0 SHA-256`24cc8dcc410fd3c82c829d476896f85eba110b1b7aac222090ae588d91a71b88` `var model = LM.LoadFromModelID("qwen3.5:2b");` Download Qwen3.5-2B-Q4_K_M.lmk1.4 GB
Alibaba Qwen 3.5 `qwen3.5:4b`	text generation, chat, code completion, math, vision, tool calling, ocr	4.54 B	256K	2.7 GB	apache-2.0
Qwen3.5 4B is a dense hybrid model using Gated Delta Networks with gated attention, delivering strong reasoning, coding, math, and tool use across 200+ languages with a native 262K context window. Architectureqwen35 Parameters4,539,265,536 Context window262,144 tokens Quantization4-bit FormatLMK Download size2.74 GB Licenseapache-2.0 SHA-256`335709a155617096c3d99a49660824923df62555029d939d77d511a625e6030a` `var model = LM.LoadFromModelID("qwen3.5:4b");` Download Qwen3.5-4B-Q4_K_M.lmk2.7 GB
Alibaba Qwen 3.5 `qwen3.5:9b`	text generation, chat, code completion, math, vision, tool calling, ocr	9.41 B	256K	5.7 GB	apache-2.0
Qwen3.5 9B is a dense hybrid model using Gated Delta Networks with gated attention, delivering state-of-the-art reasoning, coding, math, and tool use across 200+ languages with a native 262K context window. Architectureqwen35 Parameters9,409,813,744 Context window262,144 tokens Quantization4-bit FormatLMK Download size5.71 GB Licenseapache-2.0 SHA-256`9451d7307761a6e8ad41c9a37394b50663d9c593edf3e1f0b84df57d74563e3e` `var model = LM.LoadFromModelID("qwen3.5:9b");` Download Qwen3.5-9B-Q4_K_M.lmk5.7 GB
Alibaba Qwen 3.5legacy `qwen3.5:27b` superseded by `qwen3.6:27b`	text generation, chat, code completion, math, vision, tool calling, ocr	27.4 B	256K	15.9 GB	apache-2.0
Legacy model. Superseded by `qwen3.6:27b`; it remains available for compatibility, but new projects should use the replacement. Qwen3.5 27B is a dense hybrid model using Gated Delta Networks with gated attention, delivering state-of-the-art reasoning, coding, math, and tool use across 200+ languages with a native 262K context window. Architectureqwen35 Parameters27,356,728,560 Context window262,144 tokens Quantization4-bit FormatLMK Download size15.88 GB Licenseapache-2.0 SHA-256`eacae662d5096b232661f37d40577dc3244c34d99b21c557bd19b67370c11077` `var model = LM.LoadFromModelID("qwen3.5:27b");` Download qwen3.5-27b-Q4_K_M.lmk15.9 GB
Alibaba Qwen 3.5 35B-A3Blegacy `qwen3.5:35b-a3b` superseded by `qwen3.6:35b-a3b`	text generation, chat, code completion, math, vision, tool calling, ocr	35.1 B	256K	20.2 GB	apache-2.0
Legacy model. Superseded by `qwen3.6:35b-a3b`; it remains available for compatibility, but new projects should use the replacement. Qwen3.5 35B-A3B is a sparse Mixture-of-Experts model (35B total, 3B active) using Gated Delta Networks with 256 experts, delivering exceptional reasoning, coding, math, and tool use at a fraction of the compute cost of comparable dense models, with a native 262K context window. Architectureqwen35moe Parameters35,107,181,936 Context window262,144 tokens Quantization4-bit FormatLMK Download size20.18 GB Licenseapache-2.0 SHA-256`8747f18bba94761849dfc4182dae54a4867b0c80116bb3939bbe373cbda7b191` `var model = LM.LoadFromModelID("qwen3.5:35b-a3b");` Download qwen3.5-35b-a3b-Q4_K_M.lmk20.2 GB
Alibaba Qwen 3.6 `qwen3.6:27b`	text generation, chat, code completion, math, vision, tool calling, ocr	27.4 B	256K	16.4 GB	apache-2.0
Qwen3.6 27B is a dense vision-language model combining Gated DeltaNet linear attention with Gated Attention layers across 64 blocks, adding enhanced agentic coding and repository-level reasoning, thinking-context preservation across turns, and multi-token prediction over the Qwen3.5 generation, with a native 262K context window extensible to 1M tokens via YaRN. Architectureqwen35 Parameters27,356,728,560 Context window262,144 tokens Quantization4-bit FormatLMK Download size16.42 GB Licenseapache-2.0 SHA-256`90e473faf1e81f80a2ca66dcdafa3d43ac3d8d6f8770b82f825a991d19b7df76` `var model = LM.LoadFromModelID("qwen3.6:27b");` Download Qwen3.6-27B-Q4_K_M.lmk16.4 GB
Alibaba Qwen 3.6 35B-A3B `qwen3.6:35b-a3b`	text generation, chat, code completion, math, vision, tool calling, ocr	35.1 B	256K	21.6 GB	apache-2.0
Qwen3.6 35B-A3B is a sparse Mixture-of-Experts vision-language model (35B total, 3B active) with 256 experts (8 routed plus 1 shared) and a hybrid Gated DeltaNet/Gated Attention backbone over 40 blocks, adding enhanced agentic coding, thinking-context preservation, and multi-token prediction over the Qwen3.5 generation, with a native 262K context window extensible to 1M tokens via YaRN. Architectureqwen35moe Parameters35,107,181,936 Context window262,144 tokens Quantization4-bit FormatLMK Download size21.59 GB Licenseapache-2.0 SHA-256`8695a006fee5046c7573a264cc9e385b8da3ca1e3a5c0ac03b881bfc8f3b9c16` `var model = LM.LoadFromModelID("qwen3.6:35b-a3b");` Download Qwen3.6-35B-A3B-Q4_K_M.lmk21.6 GB
Alibaba Qwen QwQ `qwq`	text generation, chat, math, reasoning, tool calling	32.8 B	40K	18.5 GB	apache-2.0
QwQ is a reasoning-focused model in the Qwen series that significantly outperforms conventional instruction-tuned models on challenging tasks, with QwQ-32B demonstrating competitive performance compared to top reasoning models like DeepSeek-R1 and o1-mini. Architectureqwen2 Parameters32,763,876,352 Context window40,960 tokens Quantization4-bit FormatGGUF Download size18.49 GB Licenseapache-2.0 SHA-256`6c2c72d16bbf5b0c30ac22031e0800b982b7d5c4e4d27daa62b66ee61c565d17` `var model = LM.LoadFromModelID("qwq");` Download QwQ-32B-Q4_K_M.gguf18.5 GB
HuggingFace SmolLM3 `smollm3:3b`	text generation, chat, code completion, math	3.08 B	64K	1.8 GB	apache-2.0
SmolLM3 is a 3B parameter language model designed to push the boundaries of small models. It supports 6 languages, advanced reasoning and long context. SmolLM3 is a fully open model that offers strong performance at the 3B–4B scale. Architecturesmollm3 Parameters3,075,098,624 Context window65,536 tokens Quantization4-bit FormatGGUF Download size1.78 GB Licenseapache-2.0 SHA-256`03cc959aceff388ca737e4714a20cf4b3ef403116c9759a2a99f504dec40294e` `var model = LM.LoadFromModelID("smollm3:3b");` Download SmolLM3-3B-Q4_K_M.gguf1.8 GB
Google TranslateGemma 3 `translategemma3:4b`	text generation, chat, vision, translation	4.3 B	128K	2.6 GB	gemma
TranslateGemma is a collection of open translation models built on Gemma 3, helping people communicate across 55 languages with text and image inputs. It represents a significant step forward in open translation, no matter where they are or what device they own. Architecturegemma3 Parameters4,300,079,472 Context window131,072 tokens Quantization4-bit FormatLMK Download size2.61 GB Licensegemma SHA-256`fee7d470e1c00ac741820c23620a3338577dc8c6c7786c657d352b72951706ca` `var model = LM.LoadFromModelID("translategemma3:4b");` Download translategemma-4B-it-Q4_K_M.lmk2.6 GB
Google TranslateGemma 3 `translategemma3:12b`	text generation, chat, vision, translation	12.2 B	128K	7.0 GB	gemma
TranslateGemma is a collection of open translation models built on Gemma 3, helping people communicate across 55 languages with text and image inputs. It represents a significant step forward in open translation, no matter where they are or what device they own. Architecturegemma3 Parameters12,187,325,040 Context window131,072 tokens Quantization4-bit FormatLMK Download size7.01 GB Licensegemma SHA-256`b84df99e835b9d3b2d4d30eab222753707cbc24252a28731789d77cc3b0f97f2` `var model = LM.LoadFromModelID("translategemma3:12b");` Download translategemma-12B-it-Q4_K_M.lmk7.0 GB
U2-Net 44M `u2net`	segmentation	44 M	0	168 MB	apache-2.0
A U-square nested U-Net for salient object detection and general image segmentation; lightweight encoder–decoder with RSU blocks. Architectureu2net Parameters44,000,000 Context window0 tokens Quantization32-bit FormatLMK Download size167.85 MB Licenseapache-2.0 SHA-256`bfc5e34225e3c8d3b5c3ffd3b128c7d7e6bb17de9bde56b3a6d0654de5e73661` `var model = LM.LoadFromModelID("u2net");` Download u2-net-F32.lmk168 MB
OpenAI Whisper Base `whisper-base`	speech-to-text	73 M	1K	78 MB	mit
A balanced Whisper model delivering moderate resource use with reliable transcription accuracy. Architecturewhisper Parameters72,593,920 Context window1,500 tokens Quantization8-bit FormatGGML Download size77.98 MB Licensemit SHA-256`c577b9a86e7e048a0b7eada054f4dd79a56bbfa911fbdacf900ac5b567cbb7d9` `var model = LM.LoadFromModelID("whisper-base");` Download whisper-base-q8_0.bin78 MB
OpenAI Whisper Large Turbo V3 `whisper-large-turbo3`	speech-to-text	809 M	1K	834 MB	mit
A turbo-optimized Whisper large v3 variant for faster transcription with near-v3 accuracy. Architecturewhisper Parameters808,878,080 Context window1,500 tokens Quantization8-bit FormatGGML Download size833.69 MB Licensemit SHA-256`317eb69c11673c9de1e1f0d459b253999804ec71ac4c23c17ecf5fbe24e259a1` `var model = LM.LoadFromModelID("whisper-large-turbo3");` Download whisper-large-v3-turbo-q8_0.bin834 MB
OpenAI Whisper Large V2 `whisper-large2`	speech-to-text	1.54 B	1K	1.5 GB	mit
A high-capacity Whisper model providing excellent transcription accuracy across multilingual audio content. Architecturewhisper Parameters1,543,490,560 Context window1,500 tokens Quantization8-bit FormatGGML Download size1.54 GB Licensemit SHA-256`fef54e6d898246a65c8285bfa83bd1807e27fadf54d5d4e81754c47634737e8c` `var model = LM.LoadFromModelID("whisper-large2");` Download whisper-large-v2-q8_0.bin1.5 GB
OpenAI Whisper Large V3 `whisper-large3`	speech-to-text	1.54 B	1K	1.5 GB	mit
The largest Whisper v3 model providing state-of-the-art transcription accuracy across varied audio. Architecturewhisper Parameters1,543,490,560 Context window1,500 tokens Quantization8-bit FormatGGML Download size1.54 GB Licensemit SHA-256`37efc6b68f300ab717465685f7c3e175a66c11cf92bb3ab9912e86f4116c465e` `var model = LM.LoadFromModelID("whisper-large3");` Download whisper-large-v3-q8_0.bin1.5 GB
OpenAI Whisper Medium `whisper-medium`	speech-to-text	764 M	1K	785 MB	mit
A medium-sized Whisper model offering high-quality transcription for diverse audio scenarios. Architecturewhisper Parameters763,857,920 Context window1,500 tokens Quantization8-bit FormatGGML Download size785.23 MB Licensemit SHA-256`42a1ffcbe4167d224232443396968db4d02d4e8e87e213d3ee2e03095dea6502` `var model = LM.LoadFromModelID("whisper-medium");` Download whisper-medium-q8_0.bin785 MB
OpenAI Whisper Small `whisper-small`	speech-to-text	242 M	1K	252 MB	mit
A small Whisper model providing improved transcription fidelity while remaining efficient. Architecturewhisper Parameters241,734,912 Context window1,500 tokens Quantization8-bit FormatGGML Download size252.21 MB Licensemit SHA-256`49c8fb02b65e6049d5fa6c04f81f53b867b5ec9540406812c643f177317f779f` `var model = LM.LoadFromModelID("whisper-small");` Download whisper-small-q8_0.bin252 MB
OpenAI Whisper Tiny `whisper-tiny`	speech-to-text	38 M	1K	42 MB	mit
The smallest Whisper variant offering fast, lightweight speech-to-text transcription. Architecturewhisper Parameters37,760,640 Context window1,500 tokens Quantization8-bit FormatGGML Download size41.52 MB Licensemit SHA-256`c2085835d3f50733e2ff6e4b41ae8a2b8d8110461e18821b09a15c40c42d1cca` `var model = LM.LoadFromModelID("whisper-tiny");` Download whisper-tiny-q8_0.bin42 MB

MTEB scores are mean task scores on the multilingual MTEB leaderboard, the standard public benchmark for embedding quality; higher is better. Models without a published score show none.

Downloads are cached in the local model directory and reused across sessions; every file can be verified against the SHA-256 checksum in its row.

Legacy models remain available for compatibility and name their replacement; use the toggle above to list them.

Table of Contents

LM-Kit Model Catalog