Mistral Large 4: 1T-parameter open model in public preview

What did Mistral release?

On October 6, Mistral opened a public preview of Mistral Large 4, its largest and most capable model to date (Mistral). The API can be tried now in Mistral Studio. Mistral promises the weights by the end of October but has not announced a license yet. Internally the model is called ML4, or jokingly “le Chonk”.

NVIDIA's October 21 post on European cloud partners also refers to it as Mistral's latest frontier multimodal model (NVIDIA).

What changed from Large 3?

The design is still a granular mixture-of-experts, but much bigger. Of 1.05 trillion total parameters, 52 billion are active per token, alongside a 1.6 billion-parameter vision encoder (Mistral Docs). Mistral Large 3, released last December, had 675 billion total and 41 billion active parameters. The context window grows from 256K to 1 million tokens.

The model takes text and images as input and returns text. It supports structured outputs, function calling, document Q&A, batching and Mistral's Agents API. Mistral says it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers. The training data spans more than 160 languages, including every official EU language. Reinforcement learning is still running, so Mistral expects the preview to improve quickly.

What does Mistral report?

The figures below are Mistral's own vendor data. In coding, it reports 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA and 28.3% on Terminal-Bench 4. Its Coding Agent Index score of 49.8% puts it ahead of DeepSeek V4 Pro and Qwen3.8 Max.

Cybersecurity is the standout. On one Artificial Analysis Cyber Index test, where a model must reproduce a real vulnerability and then patch it, Mistral says ML4 scores 82%, the highest of any model, and it solves 93% of the 40 Cybench challenges. Mistral adds that several closed models score near zero on that test because they refuse the task.

For agents, Mistral reports 59.9% on AutomationBench business workflows and 1,393 Elo on AA-Briefcase, which tests long-horizon knowledge work. On the Dense 200 visual grounding test it measured 42%, against 41% for GPT-6 Astra. On Lakera's B3 prompt injection benchmark the model resists 93.3% of attacks.

What do independent tests show?

According to The Batch, Large 4 in reasoning mode scored 38 on the Artificial Analysis Intelligence Index (The Batch). That ties GPT-6 Luna and beats every US-built open model. Several cheaper Chinese open models rank higher, such as MiMo-V2.6-Pro at 46 and GLM-5.3-Flash at 42. No model built outside China and the US has scored higher on the index.

Pricing and availability

The API costs $1.36 per million input tokens, $0.14 per million cached input tokens and $4.18 per million output tokens. Large 3 is listed at $0.50 and $1.50. The 50% launch discount ran for two weeks (Mistral changelog) and has now ended. Selected cybersecurity firms, partners and government agencies can use a version with reduced moderation and expanded cyber capabilities.

As of October 21, the weights had not appeared on Mistral's Hugging Face page (Hugging Face). Mistral promises architecture details, more benchmarks and its post-training methodology alongside the weights. It also says Large 4 will be the base for its next specialized models.