Mistral has released Medium 3.5, and the early read from the AI community is that the performance-to-size ratio is genuinely impressive. The catch: it ships under a modified MIT license that prohibits commercial use.
That licensing restriction matters more than it might sound. A standard MIT license lets developers plug a model into products, client work, or internal tools without restrictions. This one does not. If you're building anything that touches a business - even a solo freelancer's automation - you're outside the permitted use. Hobbyist projects and research are fine; anything revenue-adjacent is not.
The "open weights" label still has value here. Researchers, students, and developers building non-commercial tools can download and run the model locally, which puts it in a different category from fully closed models like ChatGPT or Claude. You can inspect it, fine-tune it (adapt it on your own data), and run it on your own hardware. You just can't sell what you build with it.
For the AI developer community that's been watching Mistral's trajectory, Medium 3.5 lands between the company's lighter models and its flagship offerings. The benchmark numbers relative to parameter count - roughly, the number of internal settings that determine how the model thinks - appear to be the headline story. Smaller models that punch above their weight are practically useful because they're cheaper and faster to run.
Mistral's commercial licensing situation has been evolving. Earlier models shipped with more permissive terms. This shift toward restricted open weights puts Medium 3.5 in the same general bucket as Meta's Llama models, which also carry usage constraints despite being publicly available. Whether that's a dealbreaker depends entirely on what you're building.