Which models are open weights, and what can you do with them?
Open weights means you can download the model. The license decides what you can do with it. Here is the map, family by family.
- → Open weights means the trained parameters are published for download; it says nothing about the training data or code.
- → The license is the part that matters in practice: MIT and Apache 2.0 are permissive, custom licenses add conditions.
- → Most of the strongest open-weights models today come from Chinese labs, with Meta, Mistral, Google, and OpenAI also publishing.
The phrase "open model" gets used for everything from fully open-source projects to downloads with pages of conditions. The useful question is narrower: are the weights published, and under what license? Those two facts tell you whether you can run the model yourself, host it for others, build a product on it, or fine-tune it.
This page lists the main open-weights families and what their licenses allow. Versions change every few months; the families and their licensing habits are more stable.
The permissive ones: MIT and Apache 2.0
These are the licenses used for ordinary open-source software, and they place almost no conditions on use: run it, modify it, sell a product on it, with attribution. DeepSeek publishes under MIT. OpenAI's gpt-oss models are Apache 2.0. Most Qwen releases from Alibaba are Apache 2.0, as are many Mistral models. Zhipu's GLM models were MIT through GLM 5.2. If a model is under one of these, the licensing question is settled.
The nearly permissive ones
Moonshot's Kimi used a modified MIT license through K2. Kimi K3 has its own license: MIT-style terms, plus a requirement to display the Kimi K3 name prominently if your product has over a hundred million monthly users or $20 million in monthly revenue, and a separate agreement for AI service businesses above $20 million a year. Since GLM 5.3, Zhipu (now Z.ai) uses its own license on the same pattern: MIT-style terms, plus a security review by Z.ai before any cloud provider with more than $10 billion in yearly revenue hosts it commercially. Alibaba's Qwen 3.8 Max, the first Max model with open weights, has its own license too: a name display rule at the same large-product threshold, and a separate license for AI service businesses above $50 million a year. For almost everyone all three work like MIT. Whatever the family, check the license file of the version you want.
The custom licenses
Meta's Llama ships under its own community license. It allows commercial use, but adds conditions: a separate license for services above a large user threshold, an acceptable-use policy, and naming rules. Google's Gemma has its own terms of use with a usage policy attached. Both are widely used and the conditions rarely bite, but they are not open source in the strict sense, and lawyers notice.
What open weights does not include
Nearly none of these releases include the training data, and many do not publish the full training code. You get the finished model, a technical report of varying detail, and the license. That is enough to run and audit the model's behavior. It is not enough to reproduce it. "Open weights" is the honest term; "open source" often overstates it.
Why so many come from China
DeepSeek, Kimi, Qwen, and GLM are all Chinese labs, and they have led open-weights releases since 2025. The reasons are partly strategic and partly practical: publishing weights builds adoption fast, and it lets the rest of the world run the models on hardware and hosts of their choosing. Where the model runs, and who runs it, is a separate question from where it was made.
A recipe published in a magazine is open weights: you can cook it at home or in your restaurant. The license is the fine print under the recipe. Some say "do what you like, credit the author". Some say "not for chains above a certain size without asking". Neither gives you the test kitchen notes.
Berges AI runs only open-weights models: DeepSeek, GLM, Qwen, and Kimi, hosted on infrastructure we choose. The license for each family is on its model page. Open weights is why no vendor can pull a model out from under the product, and why you can read up on what you are talking to.
Try Berges AIRelated concepts
Things people ask
Is Llama open source?
Not by the usual definition. Its weights are open and commercial use is allowed, but the community license adds conditions that open-source licenses do not. Open weights with strings attached is the accurate description.
Can I build a commercial product on an open-weights model?
Under MIT or Apache 2.0, yes, with attribution. Under custom licenses, usually yes, subject to their conditions. Read the license file of the exact version; it is short.
Does open weights mean the model is safe or unbiased?
No. It means you can inspect and test it, which is more than a closed model offers. What you find is up to the model.
Is DeepSeek open weight?
Yes. DeepSeek publishes its model weights under the MIT license, along with detailed technical reports, which is why the same models run on many independent hosts, including Berges AI.