Mistral previews Le Chonk, a 1T open-weights model

Mistral opened a public preview of Mistral Large 4 on October 6, and it goes by a less formal name: Le Chonk. It is a 1 trillion parameter natively multimodal model with 52 billion active parameters, a hybrid MoE that unifies instruction, reasoning and agentic behavior in one model. The preview API is live today on Mistral Studio, and Mistral says the weights will be released by the end of the month. Pricing on the preview: $1.36 per million input tokens and $4.18 per million output tokens.

What the benchmarks say, and who ran them

Mistral's headline claim is that Le Chonk is the strongest open-weights model in the world on its target workloads, and it leads every open-weights model built in the US or Europe. The numbers with independent footing are the ones worth keeping:

• On the Artificial Analysis Cyber Index, an independent evaluation of how models find and fix real security flaws, it ranks in the top five globally and leads open-weights models developed outside China. On the test where a model must reproduce a real vulnerability and then patch it, Le Chonk scores 82%, the highest of any model in the index.
• On Cybench, a set of 40 security exercises, it solves 93%, one of the highest reported scores for an open-weights model.
• In a blind coding evaluation run by Surge AI, annotators rated it 3.74 out of 5, second of five models, behind Claude Opus 5 (4.22) and ahead of Kimi K3 and both GLM-5.x models.
• On Lakera's B3 AI Security Benchmark, a public test of indirect prompt injection resistance, Mistral reports it resists 93.3% of attacks.

The rest of the scoreboard is the company's own: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 59.9% on AutomationBench, and 1,393 Elo on AA-Briefcase, with Mistral claiming it also beats GPT-6 Astra on legal and finance tasks in evaluations run by the third party vals.ai. None of those are independently reproducible yet, but they are specific enough to check.

One result stands out for a different reason. Several closed frontier models, including Claude Opus 5.5 and GPT-6 Astra, score near zero on that reproduce-and-patch cyber test because they refuse the task. That is the trade the open-weights release is making: Mistral argues that defenders need the ability to prove a flaw is real without waiting on a provider's safety filters, and that losing access mid-incident is itself a risk. Until the weights drop, Mistral is red-teaming the preview with cybersecurity organizations, vetted partners and state authorities, giving them access with reduced moderation and expanded cyber capabilities.

Trained in Europe, sold on sovereignty

Le Chonk was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, and the preview is served from that same infrastructure. The model will ship in multiple regions worldwide, including a European deployment that Mistral operates end-to-end under European law. Its training data spanned more than 160 languages, including every official EU language.

The money behind this is visible: Mistral calls Le Chonk the first milestone of the roadmap funded by its 3 billion euro Series D, which it describes as the largest equity round ever raised by a European technology company, and it is scaling that European compute further over the coming months. The reinforcement learning run behind this preview is still in flight, and the company says the model has not saturated: at current scale a single training run produces roughly 33 billion tokens per day, around 16 billion of them trainable.

What it means for your setup

Three things matter if you route workloads to open-weights models. First, the price point: at $1.36/$4.18 per million tokens, this is a data point on where the open frontier sits cost-wise while it still carries the word preview. Second, the timeline: the weights arrive at the end of the month, which is when self-deployment math actually becomes possible, and Mistral says the model will be the base for a new generation of specialized Mistral models. Third, the use case: the strongest published results are in cybersecurity operations, enterprise knowledge work (spreadsheets, slides, legal and finance documents) and visual grounding, including gigapixel satellite imagery and engineering drawings.

If you are evaluating it, treat the preview API as the first test run, not the benchmark: most of the headline numbers come from the company, and the model is still improving between now and the weight release.