Mistral released the open weights for Mistral Large 4 on Thursday, putting its roughly 1-trillion-parameter flagship model on Hugging Face three weeks after a public preview. The mixture-of-experts model, nicknamed "Le Chonk," activates about 52 billion parameters per token and now scores 41% on Terminal-Bench 4.0, a 13-point jump over the preview checkpoint.
Mistral said the open-weight release continued reinforcement learning training after the preview, which shipped through its API on Oct. 6. The company describes Large 4 as a hybrid instruct-and-reasoning model that accepts text and image input, produces text and code, and is natively fluent in more than 160 languages. Large 3, its predecessor, had 675 billion total and 41 billion active parameters.
“Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors”— Arthur Mensch, Mistral co-founder and CEO
Mistral trained Large 4 from scratch on about 3,800 Nvidia Grace Blackwell GPUs in its European data centers. The model offers three reasoning settings, low, medium and high, so teams can trade depth of reasoning against token spend. High is the default, and low replaces the former "None" mode.
Key facts
- 01Mistral published open weights for Mistral Large 4 on Oct. 29, three weeks after a public preview on Oct. 6.
- 02The mixture-of-experts model has about 1 trillion total parameters with 52 billion active per token.
- 03It scored 41% on Terminal-Bench 4.0, up 13 points from the preview, and 48% on the Artificial Analysis Coding Agent Index.
- 04API pricing is $0.70 per million input tokens, $0.07 cached and $2.10 per million output tokens.
- 05The Mistral Ecosystem Open-Weight license is a modified MIT license; model-as-a-service providers need a separate commercial license.
On coding, Mistral reports Large 4 at 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task on average. Its 41% on Terminal-Bench 4.0 edges GLM-5.3 at 40% and came with 30% fewer output tokens than the preview. In Surge's blind human preference test for code, raters preferred Large 4 over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, and ranked it behind only Anthropic's Claude Opus 5.5 among the models evaluated.
Cybersecurity is the other headline use case. Mistral reports 51% on the Artificial Analysis Cyber Index, a 96.6% average refusal rate on malicious cyber requests, and 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% at preview. The company also lists 84.5% on BrowseComp and 73.3% on the FinWorkBench finance benchmark, second among open models only to Kimi K3.
Mistral lined up launch statements from Booz Allen Hamilton, Cloudflare, CrowdStrike, Hugging Face and Microsoft. Brad Medairy, president of National Cyber at Booz Allen Hamilton, said testing by the firm "demonstrated the success of the European Mistral Large 4 model against our internal benchmark." Cybersecurity experts and government authorities had been testing a version with fewer safety restrictions during the preview period.
The weights ship under the Mistral Ecosystem Open-Weight license, which Mistral describes as a modified MIT license. Anyone can use, modify, fine-tune and commercialize the model; the one condition is that businesses running a model-as-a-service offering need a separate commercial license from Mistral. Large 3 had shipped under Apache 2.0.
Hugging Face is hosting FP8 and NVFP4 checkpoints alongside community GGUF quantizations, and lists a context window of up to 1 million tokens. A quantized version runs on four Nvidia B200 or B300 GPUs, and Hugging Face says a 2-bit quant fits on a single Apple Silicon Ultra machine with 512GB of unified memory. Mistral plans smaller distilled versions, including one that runs on a laptop, but has not given dates.
Through Mistral's own API, Large 4 costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. The model is also available in Mistral Studio, Vibe Code, OpenRouter, Nvidia-hosted endpoints and an Nvidia NIM container, and on Microsoft Foundry, which Mistral says is the first cloud provider to offer it.
“Other labs have raised hundreds of billions to develop equivalent capabilities”— Arthur Mensch, Mistral co-founder and CEO
Mistral is pitching the release as a way for companies to run a frontier-class model on their own hardware and inside their own firewalls. Pierre Stock, Mistral's vice president of science, said the company wants to "create the cadence of the frontier."
Mistral has reportedly raised roughly $6 billion over three years, and the company says the reinforcement learning work behind Large 4 ran on its own compute in France, drawing about 10 megawatts.
Working on something we should cover, or seeing a story we missed? Send leads, documents, or feedback to hello@aichatdaily.com. For sensitive tips, see our secure tips page for Signal and PGP options.
Spotted an error? Email hello@aichatdaily.com with the URL and the issue, or read our full corrections policy.
