Mistral has unveiled the Mistral 3 family — and it makes a clear point: open models don’t have to be second-rate. Instead of a single flagship, there’s a whole lineup, from the data center all the way down to the edge device.
Large 3 at the Top
The top model is called Mistral Large 3. It’s a mixture-of-experts model with 675 billion total parameters, of which only around 41 billion are active per request. That’s exactly the trick that makes large MoE models affordable: you don’t pay for the full size, only for the part that actually does the work. On top of that comes a 256K-token context window — enough for seriously large codebases or long documents.
The whole thing was built in close partnership with NVIDIA. On a GB200 NVL72, Mistral says Large 3 hits 10x the performance of the previous H200 generation. That brings the cost per token down noticeably.
Nine Smaller Siblings
For many people, though, the lower end of the family will be the more interesting part. Mistral released nine smaller models, including the compact Ministral 3 suite. It’s optimized to run on NVIDIA’s edge platforms — on RTX PCs, laptops, and even Jetson boards.
In other words: a model from the same family as the frontier flagship, but small enough to run locally on your hardware. No cloud, no API bill.
What This Means
Mistral is playing a card that OpenAI and Anthropic don’t have: genuine open source across the entire range. If you need privacy, or simply want to stay independent, you get a model you can host yourself — from the server to the laptop.
Whether Large 3 really matches Opus 4.8 or GPT-5.5 will have to be shown by real benchmarks, not slides. But the strategy is clear: Mistral isn’t trying to build the single best model — it’s trying to build the best open ecosystem. And in the end, that might be the smarter lever.
Sources: