Introducing Mistral Large 4 (public preview)
C · Weak signal Vendor launch post from Mistral; benchmark results are self-reported or selected by the vendor and weights were not yet released.
Summary based on the full text·Original: about 11 minutes to read
In short
Mistral released a public preview of Mistral Large 4 on 6 October 2026, a 1-trillion-parameter multimodal mixture-of-experts model with open weights promised by the end of the month. Mistral positions it as the strongest open-weight model built outside China, trained and served in its own European datacenters. The post leans on cybersecurity, where it argues closed models refuse legitimate defensive tasks.
Key numbers
- 1 trillion / 52 billiontotal and active parameters of Mistral Large 4
- 3,800NVIDIA Grace Blackwell GPUs in Mistral's European datacenters used for training
- 61.7%DeepSWE v1.1 score (59.4% SWE-Atlas-QnA, 28.3% Terminal-Bench 4)
- $1.36 / $4.18price per million input / output tokens
As reported by the source.
English·Web page·opens mistral.ai
What it reports
- 1 trillion parameters, 52 billion active; trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs; training data spans more than 160 languages.
- Cyber: 82% on an Artificial Analysis Cyber Index test that asks a model to reproduce and patch a real vulnerability; 93% of the 40 Cybench challenges. Mistral says Claude Opus 5.5 and GPT-6 Astra score near zero on the reproduction test because they refuse.
- Coding: Coding Agent Index 49.8%; in a blind Surge AI evaluation it ranked second of five (3.74 on a 1-5 scale), behind Claude Opus 5 (4.22).
- Agentic work: 59.9% on AutomationBench (657 business workflows); 1,393 Elo on AA-Briefcase.
- Vision: 42% vs 41% for GPT-6-Astra on Dense 200.
- Safety: resists 93.3% of attacks on Lakera's B3 benchmark; KORA score 1.691 out of 2.
- RL at about 3,000 GPUs produces roughly 33 billion tokens a day, about 16 billion trainable.
- First milestone of a €3 billion Series D.
How they got it
Mix of third-party benchmarks (Artificial Analysis, vals.ai, Surge AI, Lakera) and internal evaluations chosen by Mistral. Architecture, full benchmarks and post-training details are deferred to the weights release.
Limits
- Benchmark selection and framing are the vendor's; several comparisons are against older or chosen competitors.
- The refusal-based comparison on cyber reflects safety policy, not capability.
- Preview model still in training; numbers may change.
Why it matters for companies
European firms with data-residency or sovereignty requirements get a large open-weight option they can self-host, priced at $1.36 per million input tokens through the API. For security teams, the pitch is a model that will not refuse defensive work and can run on-premise. Independent results after the weights ship are the evidence to wait for.
Summary based on the full text. Read the original for the full argument and data.Read the original on mistral.ai
- Tags
- Model releases