What Are Open-Weight Models? Licenses, Access & Limits (2026)

Open-weight models publish their trained parameters — the billions of numbers that make a neural network behave the way it does — so anyone can download, run, and adapt them. That single property is what makes self-hosting, independent benchmarking, and decentralized inference possible at all. But "downloadable" is not the same as "open source," and the license attached to the weights matters more than the download button.

Key facts

• The Llama 3.1 family spans 8B to 405B parameters under Meta's Llama Community License — commercial use allowed, with additional terms for very large deployments.
• Mistral 7B shipped in December 2023 under Apache-2.0, the permissive license that lets anyone use, modify, and sell derivatives with minimal conditions.
• DeepSeek-V3 is a 671B-parameter mixture-of-experts model with 37B active per token, released with weights and a commercial-use license on GitHub.
• The Open Source Initiative's Open Source AI Definition 1.0 requires open training data and code — a bar most "open" models do not clear.
• Distribution runs through the Hugging Face Hub, which hosts weights, tokenizers, model cards, and licenses in one versioned page per model.

How it works

Training a large language model produces a checkpoint: the full set of tuned parameters plus the tokenizer that converts text to tokens. An open-weight release publishes that checkpoint. What you typically get is the weights file, the tokenizer, a model card describing intended use and evaluation results, and a license file. What you usually do not get is the training data, the training code, or the data pipeline — the parts that would let you reproduce the model from scratch.

From download to deployment, the path is standardized. You pull the weights from the Hugging Face Hub, load them into a serving engine such as vLLM for throughput or Ollama for local use, and point your existing OpenAI-compatible client at the local endpoint. Because the weights are yours to keep, you can also quantize them down to fit smaller GPUs, fine-tune them on your own data, or distill them into smaller models — operations no closed API permits.

Licenses fall into two practical buckets. Permissive licenses, above all Apache-2.0, grant use, modification, and commercial redistribution with only attribution and license-preservation conditions; Mistral 7B and the Qwen family ship this way. Community licenses, like Meta's Llama Community License, also allow commercial use but add behavioral acceptable-use terms and extra conditions for deployments at very large scale. Read the license file in the release, not the launch blog post: the file is the contract.

The numbers

Common misconceptions

"Open weights means open source." It usually doesn't. The Open Source AI Definition 1.0 requires sufficiently detailed training data information, training code, and parameters under open terms. Nearly every major "open" release withholds the data and code, so "open-weight" is the honest label and "open-source model" is marketing unless the release clears the definition.

"Downloadable means restriction-free." Community licenses carry acceptable-use policies (no illegal activity, no large-scale disallowed content generation) and scale clauses. If you serve millions of users, the license file — not the Hugging Face download counter — decides whether you need to come back for written terms.

"Open weights are private by default." Downloading weights privately is not the same as running them privately. Weights run wherever you execute them: on your own hardware they are as private as your machine; pasted into someone else's hosted endpoint they are subject to that endpoint's logging. For the full picture, see what private inference actually requires.

"Open models are always behind." That was true for years and is now stale. The practical gap that remains is less about benchmark scores and more about everything around the model: closed vendors bundle evals, safety tuning, tooling, and uptime. Open weights give you the artifact; the operations are yours.

Current state (September 2026)

Open weights are now the default substrate of the independent AI stack: every decentralized network, every self-hosting guide, and most fine-tuning work starts from a Hub download. The active frontier questions are licensing durability (will permissive terms survive the next model generation?), data transparency (will any major lab meet the OSI definition?), and the rise of openly licensed reasoning models. DeAI News tracks new releases through the launch-sprint coverage and the provider tracker; the definitional fight with "open source" itself is covered in open weights vs open source.

FAQ

What does open-weight mean?

Open-weight means the trained parameters (weights) of a model are published for anyone to download and run. It does not automatically mean open source: most open-weight releases withhold training data and code, and carry licenses with use restrictions.

Are open-weight models free for commercial use?

Usually, but check the license. Apache-2.0 weights (Mistral 7B, Qwen3) permit commercial use with minimal conditions. Community licenses (Llama) permit it too, with extra terms — including a clause for very large deployments.

Where do you download open-weight models?

The Hugging Face Hub is the default distribution point: model pages host the weights, tokenizer, model card, and license. From there you serve them with engines like vLLM or Ollama, or fine-tune them on your own data.

Do open-weight models match closed frontier models?

The gap has narrowed substantially: leading open-weight releases now sit near frontier closed models on public benchmarks, and they win outright on deployability — you can run, modify, and host them anywhere.

Related reading

Start with what decentralized inference is to see where downloaded weights actually run, then read open weights vs open source for the licensing fight in full. To run one today, follow the gpt-oss-120b guide.