Today's concept: open-weight models - Llama, Mistral, Qwen, DeepSeek

"Open-weight" and "open-source" get used as synonyms. They aren't, and the gap is legal, not technical.

Start with what a weight is. An LLM's weights are the billions of numbers it learned in training. Think compiled binary, not source code: you get the finished artifact, not the training data or the pipeline that made it. "Open-weight" means that binary is downloadable and runnable on your own hardware instead of living behind someone else's API.

Hence the separate word. The Open Source Initiative's Open Source AI Definition (v1.0, Oct 2024) asks for the data information, the training and inference code, AND the parameters. A weights-only drop clears one of three: useful, but not open source.

Then the licenses, which vary more than people assume:

1) Llama (Meta) ships under Meta's own Community License, not an OSI-approved one. Past 700 million monthly active users on the release date, you go ask Meta. You must display "Built with Llama." And if you train a model on Llama outputs and distribute it, its name has to start with "Llama." (Llama 2 banned that outright, with no naming rule yet. Llama 3 added the naming requirement; 3.1 dropped the ban but kept the naming rule.)

2) Mistral is mostly Apache 2.0, but not all of it. Models like Devstral 2 123B carry a "Modified MIT" that cuts off once your company's monthly revenue passes $20M: then it's a commercial license or their hosted platform.

3) Qwen (Alibaba) puts most of its open-weight releases - all of Qwen3 included - under plain Apache 2.0. A few sizes are the exception: Qwen2/2.5's 72B variants and the small 2.5-3B carry a custom license with usage caps instead. Its Max flagships stay API-only.

4) DeepSeek's R1 is MIT on the code and the weights both, about as few strings as this gets. V3's code is MIT too, but its weights ship under a separate, still-permissive DeepSeek license rather than MIT outright.

The part that bites teams: the license travels through fine-tuning. DeepSeek's R1 distills are the clean example - the Qwen-based ones inherit Apache 2.0, the Llama-based ones inherit Llama's license. One release, different obligations depending on the base.

Where this sits now: open-weight went mainstream (OpenAI's gpt-oss shipped Apache 2.0), and mixture-of-experts plus 4-bit quantization is what makes self-hosting realistic: only a slice of a big model activates per token. The frontier tier keeps trending closed.

So my reflex is to open the LICENSE file, not the model card headline. Same habit as checking an npm package's license before you install it.

Quick check before you scroll: True or false: since Llama, Mistral, Qwen, and DeepSeek are all "open-weight," they all carry the same kind of permissive license.

Full breakdown + the answer: frankduah.me/learnings/2026-08-04-open-weight-models-llama-mistral-qwen-deepseek

New here? I post a bite-size AI / ML concept like this every day - follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH

#OpenWeight #AI #LLM #AIAgents #MachineLearning