Open-Weight Model Licence Tracker: Licence Terms, Thresholds and Conditions by Model

"Open weights" says nothing about what you may do with a model: the licence decides whether commercial use is allowed, whether you must show a "Built with" notice, and whether a large service needs a separate agreement first. This tracker lists the licence for 25 open-weight language model checkpoints from 18 providers, taken from each model card, and summarises the conditions written in the licence files for every provider-specific licence we could read. It is for platform teams, legal and procurement reviewers and developers choosing a model to self-host, fine-tune or offer to customers. It is a reference, not legal advice: read the licence file in the repository you download. Providers are listed alphabetically; order is never a ranking. Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Most checkpoints use a standard licence. 13 of the 25 checkpoints listed use Apache 2.0 or MIT as stated on the card, including gpt-oss-120b, Gemma 4, Granite 4.2, Hy3, DeepSeek-V4.1-Flash-1017 and GLM-5.3-Flash.
• Provider licences cluster around "Model as a Service". The Kimi K3.1-Pro, Mistral Ecosystem Open-Weight, Qwen3.8-Max and GLM-5.3 licences all define a Model as a Service business and attach a separate agreement, licence or review to it; the Muse Model License uses a similar test for offering model access as a service.
• Thresholds differ by several orders of magnitude. Stated revenue triggers run from none (Mistral Ecosystem Open-Weight License) through US$20 million (Kimi, Muse, MiniMax) and US$50 million (Qwen3.8-Max) to US$10 billion (GLM-5.3).
• The same family can carry two licences. Qwen3.8-27B is Apache 2.0 while Qwen3.8-2.4T-A95B uses the Qwen3.8-Max License; GLM-5.3-Flash is MIT while GLM-5.3 uses the GLM-5.3 License; Nemotron 3 Ultra and Super also use different licences.
• One checkpoint listed is non-commercial. EXAONE 4.5 33B uses the EXAONE AI Model License Agreement 1.2 - NC, which prohibits commercial use and limits use to research and education; K-EXAONE 2.0 from the same provider is Apache 2.0.

Status key

"Gated" means the Hugging Face page asks for an access request before download. Parameters are the card's own figures; Hugging Face metadata sometimes shows a slightly different number.

Open-weight licences by model

As of 2026-11-15. Providers alphabetical; within a provider, largest checkpoint first.

Mistral AI's models page also tags smaller open-weight models: Ministral 3 (14B, 8B, 3B), Voxtral Small and Shieldstral 1.0 as Apache 2.0, and the Timestral 1.0 time-series models as Modified MIT (source). For sizes, hardware and hosting of the largest checkpoints, see Open-Weight Frontier Models Compared.

Thresholds and triggers in provider licences

As of 2026-11-15, from the licence file in each repository. Wording is summarised; amounts and triggers are as written in the licence.

What each provider licence adds

EXAONE AI Model License Agreement 1.2 - NC (LG AI Research)

Clause 3.1 states the licensee "is expressly prohibited from using the Model, Derivatives, or Output for any commercial purposes"; use is limited to research and education (2.1a). Modified models must start their name with "EXAONE" (2.1c), the model may not be used to build competing models (3.1), and the agreement is governed by Korean law with arbitration in Seoul. Commercial use needs a separate licence from LG Management Development Institute.

GLM-5.3 License (Z.ai)

Grants free use, copying, modification, distribution and sale of the weights, code and documentation, with copyright and permission notices kept in copies. The only threshold is the US$10 billion Model as a Service security review above; questions go to the address in the licence.

Kimi K3.1-Pro License (Moonshot AI)

Uses MIT-style grant wording, including running, deploying and fine-tuning, with notices kept in copies. Section 2 covers Model as a Service, Section 3 the display requirement and Section 4 the exemptions shown in the table.

MiniMax Community License

Frames the grant around non-commercial purposes and makes commercial use (paid products, commercial hosted APIs, commercial deployment of fine-tuned versions) conditional on the notice or authorisation step. It lists prohibited uses including unlawful content, military purposes, harm to minors, harmful misinformation and discrimination.

Mistral Ecosystem Open-Weight License

Described on the card as "a modified version of the MIT license, for both commercial and non-commercial use". It defines Model as a Service as giving a third party access for inference or post-training with "meaningful control over the inputs, parameters, or training data", and sets no size threshold.

Muse Model License (Meta)

Dated 15 October 2026. Requires "Built with Muse Spark Open" in the product interface, documentation, website or settings, ships with a Usage Policy (Exhibit A), states the model is not intended for people under 18, and claims no ownership of outputs. Download is gated: the access form asks for full legal name, date of birth and organisation.

NVIDIA Nemotron Open Model License

States that works "are commercially usable" under a perpetual, royalty-free licence, claims no ownership of outputs, requires the NVIDIA notice text where the work includes a NOTICE file, ends the licence for anyone who sues alleging the work infringes, and includes export and sanctions compliance; last modified 15 December 2025. Nemotron 3 Ultra instead uses OpenMDW-1.1, which requires keeping the agreement and origin notices when distributing, ends rights for anyone who sues claiming patent or copyright infringement by the materials, and places no restrictions on outputs.

Qwen3.8-Max License (Alibaba Qwen)

Permits use, modification, distribution, hosting and fine-tuning of weights, configuration, inference code and documentation. "AI Work Assistant" means an independent product mainly for AI-assisted coding or office productivity; single-purpose tools and assistants for other domains are excluded.

Licences named but not summarised on the card

The ERNIE 5.5 Flash Base Model License Agreement v1.0 (Baidu) is named without a summary. StepFun describes its Model License 1.0 as "an Apache-2.0-Style license, with additional attribution and commercial-license conditions". The Upstage Solar card summarises derivative-model rules: prefix the name with "Solar", display "Built with Solar" and include the licence. Read each licence file before use.

What to watch next

• Licence changes between versions: families already mix licences (Qwen3.8, GLM-5.3, Nemotron 3, EXAONE), so a new checkpoint can move to a different licence.
• Model as a Service clauses: four provider licences now define MaaS; watch the definitions and thresholds if you host models for customers.
• Base and preview checkpoints: ERNIE-5.5-Flash-Base and Step-5-Preview carry their own licences; a later post-trained or final release may differ.
• Usage policies: Gemma 4 links terms of use and a prohibited use policy, and Muse models ship usage policies; these sit beside the licence and can change separately.

How to use this tracker

1. Find the checkpoint, not the family. Match the exact repository name; licences differ within families.
2. Read the licence file at the commit you download. Store a copy with the weights and record the commit hash in your model register.
3. Check your use against the trigger. Internal use, embedding in your own product and offering a model API to third parties are treated differently by most provider licences.
4. Plan attribution early. "Built with" notices and model-name display rules affect product interfaces and documentation.
5. Route through approval. Use the Open Source Licence Compliance Guide and the AI Model Change Approval Template; for the model's documentation see the Model Card and System Card Index by Provider.

Practitioner checklist

• Record licence name, version and repository commit for every open-weight checkpoint you run or fine-tune.
• Classify each deployment as internal use, embedded product feature or Model as a Service before reading the thresholds.
• Check revenue and monthly active user figures for your organisation and affiliates against any stated threshold.
• Add required "Built with" or model-name notices to product interfaces and documentation before launch.
• Confirm that non-commercial licences (such as EXAONE 4.5 NC) are limited to research and evaluation environments.
• Scan downloaded weights and pin hashes; see the Open-Weight Model Security Guide and Self-Hosted LLM Security Guide.
• Get legal review for any provider licence; this tracker summarises licence text and is updated from time to time, not on a fixed schedule.

Standards and references

• Moonshot AI: Kimi K3.1-Pro License (accessed 2026-11-15)
• Mistral AI: Mistral Ecosystem Open-Weight License (accessed 2026-11-15)
• Alibaba Qwen: Qwen3.8-Max License (accessed 2026-11-15)
• Z.ai: GLM-5.3 License (accessed 2026-11-15)
• MiniMax: MiniMax Community License (accessed 2026-11-15)
• Meta: Muse Spark Open model card and Muse Model License (accessed 2026-11-15)
• NVIDIA: Nemotron Open Model License (accessed 2026-11-15)
• LG AI Research: EXAONE AI Model License Agreement 1.2 - NC (accessed 2026-11-15)

Related NHI Mgmt Group resources: Open-Weight Frontier Models Compared · Open-Weight vs Closed Models Explained · Open Source Licence Compliance Guide · Self-Hosted LLMs vs Hosted LLM APIs · Model Card and System Card Index by Provider. This tracker summarises published licence texts and is not legal advice.