What is an open-weight AI model?

An open-weight AI model is one whose trained parameters — the billions of numerical “weights” that encode what the model has learned from its training data — are released publicly for download. Once you have the weights, you can run the model on your own hardware, fine-tune it on your own data, and inspect what the model does. What you usually cannot do is recreate the model from scratch, because the training code and training data are typically not released with the same openness.

The Open Source Initiative distinguishes this explicitly. In its 2024 framework for “open” AI, the OSI says that releasing the weights alone does not make a model open source — fully open AI would also require the training data, the training code, and the processing pipeline. Open-weight models sit in an intermediate category: you can use them, build on them, and deploy them, but you cannot fully reproduce them or audit exactly how they were trained. This matters for both commercial users (who care about what they can do with the model) and researchers (who care about what they can verify about it).

The current open-weight landscape includes several distinct families with different licenses and capabilities:

• Meta’s Llama family. Llama 3, 3.1, and 3.2 across a range of sizes (from 1B to 405B parameters). Llama’s license allows commercial use but imposes an Acceptable Use Policy, restrictions on certain use cases, and a 700M-monthly-active-user threshold above which additional terms apply.
• Mistral. Mistral 7B, Mixtral 8x7B, Mistral Large, and others. Mistral has released many models under the permissive Apache 2.0 license, which is the closest to standard open-source of any major lab.
• Google’s Gemma. Gemma 2, Gemma 3 in 2B/9B/27B parameter sizes. Google’s Gemma license allows commercial use but imposes an Acceptable Use Policy and a separate “Gemma Terms” set of restrictions.
• OpenAI’s gpt-oss. Released in 2025 under Apache 2.0 — the first OpenAI open-weight release. The Apache 2.0 license is the cleanest, but OpenAI’s gpt-oss usage policy adds behavioral restrictions on top.
• Alibaba’s Qwen, DeepSeek, Microsoft’s Phi, DBRX (Databricks), BLOOM (BigScience). Each with its own license terms and capabilities.

The license distinction is what most affects commercial use. Apache 2.0 (Mistral, gpt-oss) is the standard permissive open-source license: do almost anything, just preserve the copyright notice. Custom licenses (Llama, Gemma) add acceptable-use restrictions and threshold clauses that have to be reviewed before deploying at scale. The “open weights” label tells you the model is downloadable; the license tells you what you can do with it.

Three practical things the open-weight format enables:

• Local inference. You download the weights, set up a runtime (vLLM, llama.cpp, Ollama, or a managed service), and run the model on your own infrastructure. No API calls, no data leaving your environment.
• Fine-tuning. You take the base model and continue training it on your own data — your company’s documentation, your product catalog, your support transcripts. This produces a model much better at your specific tasks than a general-purpose API.
• Inspection. Researchers and security teams can examine the weights to study what the model has learned, audit for problematic patterns, or verify safety properties.

What open-weight is not:

• Not “free as in freedom.” The Open Source Initiative’s framing is clear: open weights ≠ open source. The training data and training process are usually closed, which means you cannot fully reproduce the model or audit the data it learned from.
• Not “free as in beer.” Running a large open-weight model requires significant compute (multiple high-end GPUs for a 70B-parameter model), and you pay for the infrastructure even if the weights themselves are free.
• Not necessarily safer. Closed-API models have content policies baked in. Open-weight models do not. If your application needs safety guardrails, you have to build them yourself or use a tool like Llama Guard.

The what-are-open-weight-ai-models-and-why-do-they-matter answer covers why companies specifically choose open-weight over closed APIs, and the why-do-companies-care answer covers the operational trade-offs.