Mistral Large 4 preview: 1T-parameter multimodal model with 49B active parameters, weights due this month
- Mistral opened a public preview of Mistral Large 4, a 1-trillion-parameter natively multimodal model with 49 billion active parameters; the weights ship by the end of October 2026.
- The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, on data spanning more than 160 languages including every official EU language.
- On the Artificial Analysis Cyber Index it ranks in the top five globally, and it scores 82% on the index test that reproduces and patches a real vulnerability, the highest of any model, plus 93% on Cybench.
- Claude Opus 5.5 and GPT-6 Astra score near zero on that vulnerability test because they refuse the task; Mistral argues open weights and self-deployment matter when provider refusals block legitimate security work.
- The preview API runs on Mistral Studio while Mistral red-teams the model with cybersecurity partners and state authorities who get reduced moderation and expanded cyber capabilities, using the same training and RL environment it sells through Mistral Forge.
Hacker News opinions
Woah, this seems like a big deal, assuming the benchmarks are as good as claimed. Mistral is proving me wrong here and I'm not mad about it.
They claim parity with GLM 5.3 on DeepSWE, and it looks like they're doing 50% off to stay price competitive with DeepSeek Flash V4.1.
Benchmarks are better than I expected, and they probably got there without distilling from the Chinese open weights.
Why wouldn't they distill locally running Chinese open weights? Everyone is quietly doing it, that's not a secret.
The vision numbers are the impressive part. If the vision model is really on par with Astra that's best in the world, and the cyber scores beat every Chinese model, so this is a solid defender model.
Does Mistral ever advance the state of the art on any dimension? Sovereign AI is a joke when a post-trained open weight model from an American or Chinese lab does the same thing, and they burn most of their GPU hours reproducing a last-gen pretraining run.
Because Europeans want a strong open model that isn't owned by American trillion-dollar companies or built by the Chinese, and cybersecurity is exactly where that matters.
By the benchmarks it's about a year behind, and on the vals.ai index more like three generations. Still good enough that plenty of people would use it.
What's interesting is that AI so far isn't winner-take-all. DeepSeek basically publishes instruction manuals in paper form, which is what makes catching up possible at all.
The dynamics look more like cloud computing than web search. That said, everyone I know doing serious programming is on Claude or Codex.
Training from scratch on 3,800 Blackwell GPUs in your own datacenters is capital intensive. That's not something a random team can replicate.
I'd rather see Mistral compete with Qwen3.8-Flash-Next, a 120B class model I can run locally for real coding. I want one of those but made in the EU.
We haven't hit recursive self-improvement yet. Once somebody does, the runaway scenario actually happens.
The previous Mistral is barely in the top 50 on arena.ai.