Mistral today released the downloadable open weights of Mistral Large 4, the trillion-parameter flagship it previewed on October 6 and nicknamed 'Le Chonk' after its cat mascot.

The final model is natively multimodal, with 1.05 trillion total parameters and about 52 billion active per token, and it spent the three weeks since the preview in continued reinforcement learning plus customer and security-partner testing.

Mistral says the gains show 'no signs of saturation,' and VP of science Pierre Stock calls the release 'a pivotal moment in the global AI race' and 'a turning point for Europe.'

Weights are on Hugging Face; the model also runs via Mistral's API, OpenRouter, Microsoft Foundry, and Nvidia-hosted endpoints.

Open-weight AI has been led almost entirely by Chinese labs like DeepSeek and Zhipu's GLM family, with Western releases voluntary and one-off.

A European model now claims the best open agentic coding performance is the surprise: Mistral reports Large 4 climbed from 28% to 41% on Terminal-Bench 4.0, tying GLM-5.3 as the top coding-agent model, and says it tops Artificial Analysis' Coding Agent Index among open models.

Stock framed it bluntly as the West finally fielding its own frontier open weights, arguing Chinese open models 'can be threat vectors' and cannot be fully trusted by enterprises and governments.

Artificial Analysis' independent October 29 assessment confirms a real but narrower win: at high reasoning, Large 4 scores 26 on Intelligence Index v5, up from 25, one point above Muse Spark Open and the highest-scoring weighted model developed outside China — but still six points behind open-weight leaders DeepSeek V4.1 Flash 1017 and GLM-5.3 at maximum reasoning.

On the Coding Agent Index, VentureBeat reports Artificial Analysis ranks ML4 second at 48, one point behind GLM-5.3, while cutting tokens per task from 15.9 million to 8.5 million — roughly 47% fewer.

VentureBeat also notes Mistral's own Terminal-Bench 41% differs from AA's 25% in a separate evaluation context, so the two numbers should not be conflated.

The biggest change since the preview is economic rather than a capability leap.

Artificial Analysis measures cost per Intelligence Index task at $2.06, down 63% from $5.54 for the preview, on new API rates of $0.70 per million input tokens and $2.10 per million output tokens — roughly half the preview's pricing.

Mistral trained on 3,800 Nvidia Grace Blackwell GPUs in its European data centers, and CEO Arthur Mensch says the company has taken the global lead in compute efficiency, building frontier capability on roughly $6 billion raised over three years against rivals' hundreds of billions.

Three reasoning modes — low, medium, high — let teams match depth to task and control token spend.

The pitch to enterprises is control. Open weights can run inside a customer's own firewalls and cannot be switched off remotely — a fear crystallized in June when a U.S. export-control directive forced Anthropic to suspend access to Fable 5 and Mythos 5 for nearly three weeks, per Fortune.

Stock adds that closed models rate-limited Hugging Face's own defenders during July's sandbox-escape incident, pushing the team onto a Chinese open model.

The trade-off is real: Sophos' Rafe Pilling tells WIRED it is 'clearly a dual-use technology,' since the same freely downloadable weights can be customized by threat actors, and careful monitoring matters more than country of origin.

Mistral calls Large 4 the start of a new generation spanning all modalities, with more models coming in the months after the recent Voxtral 3 voice launch — and Nvidia's same-day Nemotron 3.5 Transcribe open-weight release shows the open-weights space is crowded and accelerating.

The license matters for adopters: VentureBeat reports ML4 ships under a modified MIT license requiring a separate commercial deal only for model-as-a-service operators, with internal use exempt.

Policy pressure is also building, as a new Manhattan Institute memo proposes U.S. antitrust safe harbor for frontier labs in exchange for scheduled, independently tested open-weight releases — an American answer to the gap Mistral just narrowed for Europe.