The BigScience OpenRAIL-M license was published for the BLOOM model. It allows broad use and redistribution of weights but attaches a list of prohibited use cases that must be passed on to downstream users, establishing the behavioral-use license pattern for model weights.
Open-Weights Licensing
Also known as: model licenses · open-weight licenses · OpenRAIL
Plain English
The legal terms that decide what you may do with downloadable AI model weights. Some weights use standard licenses such as Apache 2.0 or MIT, while others use custom terms that limit commercial use, user counts, or behavior.
Technical Overview
A released set of weights is governed by a license that may differ from the license on the training and inference code. Terms range from permissive software licenses (Apache 2.0, MIT) to behavioral-use licenses (OpenRAIL), to custom community licenses with user-count thresholds, acceptable-use policies, or non-commercial limits. Weights from Llama, FLUX, and Tencent Hunyuan ship under custom terms, so they are open-weight rather than open source; the Open Source Initiative's definition additionally asks for detailed training-data information. License text travels with Checkpoints on hubs such as Hugging Face, so check it per model version.
Timeline (8)
2022
Milestone2023
ReleaseMeta released Llama 2 under the Llama 2 Community License Agreement. It permits commercial use, but a licensee whose products exceeded 700 million monthly active users in the preceding calendar month must request a separate license from Meta. Because of this custom limit, the weights are described as open-weight rather than open source.
- Release
Mistral AI released Mistral 7B under the Apache 2.0 license, a standard permissive software license with no user-count thresholds or use restrictions in the license text. This set a reference point for unrestricted commercial use of weights.
2024
ReleaseBlack Forest Labs launched FLUX.1 with three tiers. FLUX.1 [dev] weights are downloadable under a non-commercial license, while FLUX.1 [schnell] uses Apache 2.0. The split shows how one model family can carry different license terms per variant.
- Definition
The Open Source Initiative published Open Source AI Definition 1.0. It requires freedom to use, study, modify, and share an AI system, and calls for the model parameters, code, and sufficiently detailed information about training data. Under this definition, weights released with use restrictions or without data information do not qualify as open source AI.
2025
ReleaseDeepSeek released DeepSeek-R1 with weights under the MIT License, one of the most permissive terms used for a frontier-scale reasoning model, explicitly allowing distillation and commercial use.
- Release
OpenAI released gpt-oss-120b and gpt-oss-20b with weights under Apache 2.0, its first open-weight language models since GPT-2. Combined with Apache 2.0 releases from Qwen and Mistral AI, permissive licensing became common among major open-weight LLMs.
2026
MilestoneGoogle released Gemma 4 under Apache 2.0. Earlier Gemma versions used custom Gemma Terms of Use with an incorporated Prohibited Use Policy, so the change replaced bespoke terms with a standard license for the family.
See also (18)
Open full graphOpen-Weights Licensing
13 more connections
- Allen Institute for AIIncoming · Enables
- FLUXOutgoing · MentionsIncoming · Uses
- QwenOutgoing · MentionsIncoming · Uses
- Stability AIIncoming · Uses
- Video Generation Models (Sora, Veo, Kling, Wan)Incoming · Uses
- WanIncoming · Uses
- CohereIncoming · Mentions
- DeepSeekOutgoing · Mentions
- GemmaOutgoing · Mentions
- Google DeepMindOutgoing · Mentions
- Meta AIOutgoing · Mentions
- OpenAIOutgoing · Mentions
- Tencent HunyuanOutgoing · Mentions