Mistral Large 4 Cuts Cost Per Task 63% and Opens Its Weights to All

Quick Facts

• Mistral Large 4, nicknamed 'Le Chonk,' went live on Hugging Face on Oct. 29 with 1.05 trillion total parameters and 52 billion active parameters per token.
• The model costs $0.70/$2.10 per million input/output tokens, roughly 50% less than its October 6 preview price of $1.36/$4.18.
• Three weeks of additional reinforcement learning pushed its Terminal Bench 4.0 score from 28% to 41% accuracy since the preview launch.

Mistral AI released the open weights for Mistral Large 4, its most capable model to date, on Oct. 29. The release follows a three-week preview period that began Oct. 6, during which the model was available only through Mistral's API.

The weights are free to download on Hugging Face under the model name Mistral-Large-4-1T-A52B. The files ship in FP8 format across 54 files, and the model can run on as few as four Nvidia B200 or B300 GPUs.

Architecture

Large 4 is a Mixture-of-Experts model with 128 routed experts, four selected per token, plus one shared expert. It was trained on 3,800 Nvidia Grace Blackwell GPUs inside Mistral's European datacenters. The model accepts native image input and supports a one million token context window.

Mistral used the nickname 'Le Chonk' in its official launch post, alongside the formal abbreviation ML4, as a direct reference to the model's size.

Benchmarks and Cost

On Cognition's private FrontierCode 2.0 benchmark, which judges whether a repository maintainer would accept an AI-generated code change, Large 4 scored 19.7%, placing it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max among open models outside China. It posted a 61.7% on DeepSWE v1.1 and 59.9% on AutomationBench.

The economic shift is substantial. Artificial Analysis data puts the updated model at $2.06 per Intelligence Index task, down from $5.54 for the preview, a 63% drop. At comparable performance levels, it also requires 47% fewer tokens per task than GLM-5.3.

Chinese models still lead on the overall Artificial Analysis Intelligence Index. DeepSeek V4.1 Flash and GLM-5.3 each scored 32 points; Large 4 scored 26. DeepSeek and MiMo also achieve higher overall scores at lower measured task costs than Large 4.

On hallucination, Large 4's rate fell to 28% on the AA-Omniscience v1.1 benchmark, down from 41% in preview. In a Surge AI blind coding evaluation, it scored 3.74 out of 5, placing second behind Claude Opus 5, which led at 4.22.

Cybersecurity Focus

Mistral has built a commercial partnership around Large 4's security performance. The model resists 97.1% of attacks on the B3 Agent Security Benchmark, up from 93.3% in preview. It previously scored 82% on a test requiring models to reproduce and patch real software vulnerabilities, and solved 93% of challenges in Cybench.

Orange Cyberdefense, one of Europe's largest cybersecurity firms, has adopted the model as a sovereign AI solution for European enterprises. Mistral CTO Timothée Lacroix said the partnership targets organizations seeking AI built and operated in Europe.

Reasoning Modes

Government and enterprise testers pushed Mistral to add more control over how long the model thinks before responding. The company responded by adding low and medium reasoning modes alongside the existing high mode, and made the high mode 50% more efficient.

Teams can select reasoning depth per task without rewriting prompts, matching quick queries to the low setting and complex multi-step work to high.

Open Weights as Strategy

Mistral CEO Arthur Mensch framed the open release as a check on concentration. 'Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors,' he said. VP of Science Pierre Stock added that once weights circulate online, access cannot be withdrawn, an argument aimed directly at enterprise buyers weighing vendor lock-in risk.

For software companies evaluating frontier models, the pricing cut and open weights lower the barrier to self-hosting a capable coding agent at a cost that now competes more closely with commercial API alternatives.