Apache 2.0
OSI-approved permissive license with broad use, modification and redistribution rights plus an express patent grant.
Compare what open-weight model licenses actually allow — and what they still require.
A searchable reference across all 32 Gold Passports. Commercial use, modification, redistribution, hosted-service treatment, attribution, scale thresholds and incorporated usage policies are kept as separate fields instead of compressed into an “open / closed” badge.
The OpenWeightModels License Explorer is a model-by-model comparison of the legal terms attached to downloadable AI weights. It does not infer rights from the words “open”, “community”, “research” or “open weight”.
For each checkpoint, the explorer records the exact license or terms document, commercial-use status, modification rights, redistribution conditions and material special clauses. This matters because two models with equally downloadable weights can create very different obligations for a company building an API, redistributing a derivative, fine-tuning a model or serving millions of users.
The explorer also distinguishes the model license from a separate acceptable-use or safety policy. Where a license incorporates such a policy by reference — as Meta's Llama Community agreements do — that relationship is surfaced. Where a publisher releases model weights under Apache 2.0 and separately publishes a usage policy, the site does not rewrite Apache 2.0 to pretend the two documents are the same instrument.
The cards group related terms for navigation. The comparison table below preserves the exact per-model license name and special conditions.
OSI-approved permissive license with broad use, modification and redistribution rights plus an express patent grant.
Short OSI-approved permissive license allowing use, modification, sublicensing and sale while preserving the copyright/permission notice.
Google model-specific terms. Distribution is defined broadly and includes making functionality available through hosted services.
Custom Meta agreements with attribution, notice, AUP and large-platform conditions; Llama 4 multimodal adds an EU policy restriction.
Kimi K2 keeps the MIT grant but adds a scale-triggered user-interface branding condition.
Custom Qwen agreement with a 100M-MAU commercial threshold and model-attribution conditions for certain derived AI models.
Search by model or developer, filter by license family and narrow by commercial-use classification.
| Model | License | Commercial | Modification | Redistribution | Material condition | Source |
|---|---|---|---|---|---|---|
| DeepSeek-R1 DeepSeek | MIT | Allowed | Allowed | Allowed with copyright and permission notice | Short permissive license. It does not itself impose a use-field restriction; application law, provider terms and separate policies still apply. | License ↗ |
| DeepSeek-V3.1 DeepSeek | MIT | Allowed | Allowed | Allowed with copyright and permission notice | Short permissive license. It does not itself impose a use-field restriction; application law, provider terms and separate policies still apply. | License ↗ |
| Devstral Small 2505 Mistral AI | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Gemma 3 12B IT Google DeepMind | Gemma Terms | Conditional | Allowed subject to terms | Conditional; distribution obligations apply | The Gemma Terms define distribution broadly enough to include making Gemma or derivatives available through hosted services. Recipients, notices, modified-file marking and prohibited-use obligations should be reviewed before redistribution. | License ↗ |
| Gemma 3 27B IT Google DeepMind | Gemma Terms | Conditional | Allowed subject to terms | Conditional; distribution obligations apply | The Gemma Terms define distribution broadly enough to include making Gemma or derivatives available through hosted services. Recipients, notices, modified-file marking and prohibited-use obligations should be reviewed before redistribution. | License ↗ |
| Gemma 3 4B IT Google DeepMind | Gemma Terms | Conditional | Allowed subject to terms | Conditional; distribution obligations apply | The Gemma Terms define distribution broadly enough to include making Gemma or derivatives available through hosted services. Recipients, notices, modified-file marking and prohibited-use obligations should be reviewed before redistribution. | License ↗ |
| Gemma 3n E4B IT Google DeepMind | Gemma Terms | Conditional | Allowed subject to terms | Conditional; distribution obligations apply | The Gemma Terms define distribution broadly enough to include making Gemma or derivatives available through hosted services. Recipients, notices, modified-file marking and prohibited-use obligations should be reviewed before redistribution. | License ↗ |
| GLM-4.5 Z.ai | MIT | Allowed | Allowed | Allowed with copyright and permission notice | Short permissive license. It does not itself impose a use-field restriction; application law, provider terms and separate policies still apply. | License ↗ |
| GLM-4.5 Air Z.ai | MIT | Allowed | Allowed | Allowed with copyright and permission notice | Short permissive license. It does not itself impose a use-field restriction; application law, provider terms and separate policies still apply. | License ↗ |
| gpt-oss-120b OpenAI | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| gpt-oss-20b OpenAI | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Granite 3.3 8B Instruct IBM | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Kimi K2 Instruct Moonshot AI | Modified MIT | Allowed with special condition | Allowed | Allowed with MIT notice requirements | If a commercial product/service exceeds either 100M monthly active users or $20M monthly revenue, the published modified license requires prominent “Kimi K2” display in the product/service user interface. | License ↗ |
| Llama 3.3 70B Instruct Meta | Llama 3.3 Community | Conditional | Allowed subject to terms | Conditional with Llama attribution/notice requirements | Commercial and research use is intended, but the Community License and Acceptable Use Policy remain part of the deployment decision. Review the current official agreement for scale, attribution and redistribution conditions. | License ↗ |
| Llama 4 Maverick Meta | Llama 4 Community | conditional | allowed subject to terms | conditional | If licensee/affiliate products or services exceeded 700M MAU in the calendar month preceding the Llama 4 release date, a separate license from Meta is required before exercising rights. | License ↗ |
| Llama 4 Scout Meta | Llama 4 Community | Conditional | Allowed subject to terms | Conditional with attribution/notice requirements | Includes the 700M-MAU special licensing condition; the incorporated Llama 4 AUP contains a specific EU developer restriction for multimodal Llama 4 models while excluding end users of products/services from that particular restriction. | License ↗ |
| Magistral Small 2506 Mistral AI | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Mathstral 7B v0.1 Mistral AI | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Mistral Nemo Instruct 2407 Mistral AI / NVIDIA | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Mistral Small 3.1 24B Instruct Mistral AI | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| OLMo 2 13B Instruct Ai2 | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| OLMo 2 32B Instruct Ai2 | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Phi-4 Microsoft | MIT | Allowed | Allowed | Allowed with copyright and permission notice | Short permissive license. It does not itself impose a use-field restriction; application law, provider terms and separate policies still apply. | License ↗ |
| Phi-4 Mini Instruct Microsoft | MIT | Allowed | Allowed | Allowed with copyright and permission notice | Short permissive license. It does not itself impose a use-field restriction; application law, provider terms and separate policies still apply. | License ↗ |
| Phi-4 Multimodal Instruct Microsoft | MIT | Allowed | Allowed | Allowed with copyright and permission notice | Short permissive license. It does not itself impose a use-field restriction; application law, provider terms and separate policies still apply. | License ↗ |
| Qwen2.5-Coder-32B-Instruct Qwen / Alibaba | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Qwen2.5-VL-72B-Instruct Qwen / Alibaba | Qwen License | Conditional | Allowed subject to terms | Conditional with agreement and Notice requirements | The official Qwen license requires a separate license for specified commercial products/services above 100M monthly active users and requires “Built with Qwen” or “Improved using Qwen” attribution when Qwen materials/outputs are used to create or improve a distributed AI model. | License ↗ |
| Qwen3-235B-A22B Qwen / Alibaba | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Qwen3-30B-A3B Qwen / Alibaba | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Qwen3-32B Qwen / Alibaba | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| Qwen3-Coder-480B-A35B-Instruct Qwen / Alibaba | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
| SmolLM3 3B Hugging Face | Apache 2.0 | Allowed | Allowed | Allowed with license/notice conditions | Includes an express patent grant. Model-specific acceptable-use policies, if separately published, are not automatically part of Apache 2.0. | License ↗ |
A one-word license label is insufficient for model governance. These dimensions are stored independently in each Passport.
Can an organization use the weights in a revenue-generating product or service, and does a scale threshold change that right?
Can the model be fine-tuned, merged, adapted or otherwise modified under the published terms?
What must accompany redistributed weights or derivatives — license text, notices, attribution or naming?
Does exposing functionality over an API/web service count as distribution under the exact terms?
Are there naming or attribution conditions when outputs or materials are used to train another AI model?
Is an acceptable-use document separate, optional, publisher guidance, or explicitly incorporated into the agreement?
Do user-count, revenue or regional conditions alter eligibility or trigger additional obligations?
Licenses are attached to a model release, not assumed from a neighboring family member. Qwen2.5-VL and Qwen2.5-Coder are a useful example.
Apache 2.0 and MIT are OSI-approved software licenses, but calling an entire AI system “Open Source AI” involves more than seeing one of those labels on a weight repository.
The Open Source Initiative's Open Source AI Definition 1.0 describes freedoms to use, study, modify and share the system and also addresses access to the preferred form for modification. OpenWeightModels therefore avoids turning a software-license field into a universal “open-source AI” badge.
Conversely, a custom model license does not make the weights “closed” if the trained parameters are publicly obtainable. The registry uses open weight as the access category, then separately documents the exact rights.
These examples show why the explorer stores clauses, not only SPDX-like names.
Meta's Llama 3.3 and Llama 4 agreements require “Built with Llama” attribution for specified distribution scenarios, include Notice-file requirements and contain a 700M-MAU additional commercial condition. Llama 4 also incorporates an AUP with a multimodal EU developer restriction.
Google's Gemma Terms define “Distribution” broadly enough to include making Gemma or derivatives available as a Hosted Service. That makes hosted deployment a legal dimension worth surfacing, not only an infrastructure choice.
The Qwen License includes a separate authorization condition above 100M monthly active users for specified commercial use and requires “Built with Qwen” or “Improved using Qwen” in certain derived-model scenarios.
Moonshot's Modified MIT license preserves broad MIT-style permissions but adds prominent “Kimi K2” UI branding above either 100M MAU or $20M monthly revenue.
Apache 2.0 grants broad rights and includes an express patent license, while redistribution retains license/NOTICE obligations where applicable. A separate publisher usage policy should be recorded separately unless legally incorporated.
The MIT License grants broad use, modification, distribution, sublicensing and sale rights subject to preserving the copyright and permission notice in copies or substantial portions.
The Explorer is a comparison layer, not a substitute for current legal text.
No universal rule exists. In this 32-model registry many checkpoints use Apache 2.0 or MIT, while Llama, Gemma, Qwen2.5-VL and Kimi K2 use terms with additional conditions. Read the exact model release.
Both are OSI-approved permissive licenses. They differ in text and obligations, including Apache 2.0's express patent provisions. OpenWeightModels does not rank one as an overall openness winner.
Not in this registry's terminology. If trained weights are publicly obtainable they can qualify as open-weight here, while the custom terms are documented separately.
No. The explorer is a technical comparison and provenance layer. Current official legal text and organization-specific legal review remain authoritative for a particular use.
Move from legal permissions to self-hosting, serving runtimes, hardware class and managed deployment paths.