LM-Kit OneDocs2026.8.10lm-kit.com
Evaluate

Licensing and What You May Build

What procurement and legal need to sign off before this server carries production traffic. Two distinct layers, cleared separately: the product license that governs operating LM-Kit One, and the upstream license that rides every model you run on it. This page summarizes for orientation; the EULA is canonical and prevails wherever the two differ, and nothing here is legal advice.

1Two layers, one review#

  • The product license. One EULA covers LM-Kit One and the LM-Kit.NET SDK it embeds. It decides who may run the server in production, on what tier, and what kind of product may be built on it. For Professional customers, commercial terms (including pricing) live in the per-customer order, not in the EULA itself.
  • The model weights. The product license grants nothing about models. Every model in the catalog carries its OWN upstream license, published by the model's vendor, and that license governs what you may ship with those weights.

A review that clears only one layer is incomplete. Clear both.

2The product license#

There are no license keys, no activation, and no runtime license checks: the tier that applies is determined solely by the criteria in the EULA, assessed at the time of use.

  • Evaluation and development are free for everyone, with no time limit and no restriction on organization size.
  • Production use and redistribution are free for a Licensee that satisfies all of the following: annual gross revenue below $1,000,000 USD, 10 or fewer employees, and no more than $3,000,000 USD received in aggregate from outside capital (venture capital and private equity included). Personal use, education, registered nonprofits, and open source projects get the same rights free of charge regardless of size. Employees, customers, and end users of an application built with the software do not count toward the employee figure.
  • The free tier requires attribution. Production use under the free grant requires a clearly visible acknowledgement of LM-Kit on the product's website or in its documentation.
  • The criteria aggregate. Eligibility is assessed for the Licensee together with its parent, subsidiary, and affiliated entities.

A Licensee that does not satisfy those criteria requires a Professional License for production use and redistribution.

3Who the license is assessed against#

The tier follows whoever benefits from production use, not whoever writes the code. The EULA is explicit on three points that trip up agencies and platform teams:

  • The beneficiary is the Licensee. Production use is assessed against the entity that owns or commissions the application, or on whose behalf the software is operated. A contractor, agency, or freelancer building or deploying for a client does not extend its own eligibility to that client: the client must itself qualify for the free tier or hold a Professional License.
  • Operating LM-Kit One is production use by the entity the server runs for, regardless of who installed it. An IT provider standing up a server for a customer puts that customer, not the provider, on the hook for the license.
  • Redistribution covers end users of your own products. It does not cover custom development delivered to a specific client; that client is a Licensee in its own right.

4The Professional License#

  • The Professional License is subscription-based and must be renewed every year to continue using the software. There is no perpetual fallback in the EULA: a lapsed subscription is a lapsed license.
  • The EULA itself contains no per-server or per-seat metering: the tier turns on the Licensee criteria alone. Whether an order scopes the deployment (servers, seats, environments) is a matter for that order.
  • Commercial terms, including pricing, are set out in the applicable order, placed directly with LM-Kit or through an authorized reseller. The order and the EULA together are the agreement your legal team reviews.
  • The EULA also reserves the Licensor's right to cancel the contract annually, with at least 60 days' notice.

5What you may build#

The EULA permits embedding the software as an integrated part of an application that adds substantial functionality of its own. It prohibits building a product aimed at third-party developers or whose primary purpose is to expose the software's capabilities: an SDK, library, component, framework, or a general-purpose AI, inference, OCR, document-processing, or RAG API or server. Products in that category require a separate agreement with LM-Kit.

The practical line: an invoice-automation product that uses this server for extraction is an application adding substantial functionality. Reselling this server's endpoints as your own inference or OCR API is not; contact LM-Kit before building that.

6Model weight licenses#

The license attaches per model, not per vendor, and can differ within one vendor's lineup. Across the current catalog, most models carry a permissive open-source license: the Qwen 3.x generations (qwen3.5, qwen3.6, qwen3.8, qwen3-coder:30b-a3b, the qwen3-embedding family), the gemma4 family, gptoss:20b, and muse-glimmer:30b are Apache-2.0; the GLM models (glm4.7-flash, glm-4.6v-flash, glm-ocr), the whisper family, bge-m3, and harrier-oss:0.6b are MIT. A minority carry family-specific vendor terms instead: embeddinggemma-300m, for example, is governed by Google's Gemma Terms of Use even though the gemma4 generation is Apache-2.0.

How to check before shipping:

  • Catalog models carry their license identifier on the model card. Query it over the Ollama-dialect show endpoint (POST /api/show, see API Compatibility), then read the named license on the vendor's page for the terms themselves.
  • Imported models are your responsibility. The import inspects the file and records the license embedded in its metadata when one is present; when none is present, no license is recorded and none is shown, and the import form lets you record the license you verified yourself. The value shown beside an imported model in the Models section is only as good as that verification.
  • Fine-tuned artifacts are derived from their base model, so obligations under the base model's license (attribution notices, and use restrictions where the family imposes them) travel with the derived weights. Verify the base model's license before shipping a fine-tune (see Fine-tuning).

7A model-approval workflow#

Organizations that must control which weights run in production can enforce the decision rather than document it:

  • Pin the server to the catalog. The operator can disable custom model imports entirely (see Choosing and Operating Models), so nothing serves that legal has not had the chance to review against a known license.
  • Restrict who may acquire weights. Getting a model onto the server is a gated act: the Models section sits on the admin plane, and API-side pulls over the Ollama dialect answer only to callers your key policy admits (see The Security Model). The approval decision is enforceable through who holds admin access and keys, not only documented.
  • Record the approved lineup. The Models section lists what is on disk; the license identifiers above give legal a per-model answer rather than a per-vendor guess.
  • Provenance is on record. Every import is recorded with its origin, who imported it, when, and a SHA-256 of the artifact's bytes computed at import; each catalog card carries the expected SHA-256 of its file. When legal asks where the weights came from, the answer exists (mechanics in Choosing and Operating Models).
  • Review at adoption, not at incident. A model's license is fixed when you pull it; one check per model-adoption decision is the whole cost.

8Stated plainly#

  • Two layers: the product EULA and each model's own license. Both must clear before production.
  • The free tier is real (thresholds in section 2), requires visible attribution, and is assessed against the entity the server runs for, not whoever installed it.
  • The Professional License renews annually; commercial terms live in the order.
  • Embedding in a value-added application is permitted; reselling general-purpose AI, inference, OCR, document-processing, or RAG APIs requires a separate agreement.
  • The EULA page is the single source of trust; this guide orients, it does not bind.