Mistral Large 4 Preview Released: 1 Trillion Parameters on Grace Blackwell

Mistral has launched a public API preview of Mistral Large 4, a natively multimodal 1-trillion-parameter model with 49 billion active parameters. Unofficially dubbed ML4 and very officially named “le Chonk,” the model represents a significant capability jump for the European company. While the API preview is available immediately, the open weights are slated for release by the end of October.

Mistral announced the launch, emphasizing that ML4 pushes the frontier of open-weight performance. The model was trained from scratch in Mistral’s European datacenters on a cluster of 3,800 NVIDIA Grace Blackwell GPUs. The public preview is currently being served on that same infrastructure, marking a major milestone in their long-term infrastructure investment.

A significant theme of the release is its focus on critical enterprise workloads, particularly cybersecurity. Mistral asserts that ML4 is one of the world’s strongest AI models for security operations, leading open-weight models developed outside of China. It achieves high marks on the Artificial Analysis Cyber Index. Mistral also reports it solved 93% of challenges in Cybench.

Mistral argues that open weights are essential for legitimate vulnerability research and incident response. On a specific Artificial Analysis test asking models to reproduce and patch a real open-source vulnerability, Mistral claims Claude Opus 5.5 and GPT-6 Astra score near zero because they refuse the task, whereas ML4 achieves 82%—the highest of any model they tested. ML4 pairs these cyber capabilities with the autonomy of self-deployment, avoiding provider-level refusals that can hinder defensive operations. In internal testing, the model has proven useful for analyzing malware, prioritizing vulnerabilities, and writing detection rules.

Beyond cybersecurity, ML4 is designed for agentic coding and complex terminal workflows. Mistral reports that it excels across software engineering and repository understanding, scoring 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, and 28.3% on Terminal-Bench 4. According to the company, its combined Coding Agent Index score of 49.8% places it ahead of competitors like DeepSeek V4 Pro 0813 and Qwen3.8 Max. Mistral also states the model runs general-purpose agents for business workflows effectively; on AutomationBench, evaluating tasks across Gmail, Google Sheets, Slack, and Salesforce, it scores 59.9%.

ML4 incorporates vision capabilities, allowing it to reason powerfully across complex documents, charts, and natural images. This multimodal integration extends to visual grounding, useful in engineering and geospatial analysis. The model also boasts strong scientific capabilities, with applications spanning applied mathematics, physics, and chemistry. Mistral highlights its proficiency in agentic coding for scientific tasks, generating full Hartree–Fock simulations in one shot.

The broader AI community has quickly begun testing the API preview. Simon Willison’s initial impressions highlight that Mistral is back in the game. Willison notes that the Mistral API for ML4 supports two reasoning levels: “none” and “high”. In his testing, generating an SVG of a pelican riding a bicycle, the “high” reasoning level produced a substantially better result while surprisingly using fewer output tokens (2,717) compared to the “none” setting (3,275). He notes that ML4 scores 38 on the Artificial Analysis leaderboard, placing it just behind the 552-billion-parameter DeepSeek 4.1 Flash, marking a massive improvement over last December’s Mistral Large 3.

Willison also shared a comparative test on Hacker News, prompting various models to “Generate an SVG of an armadillo in fishnet tights jaywalking on Mars” to see how they handle saturated benchmarks and complex, unusual requests. This highlights the ongoing community effort to probe the edges of frontier models and evaluate their creative and reasoning capabilities in practice.

The release also underscores Mistral’s ongoing post-training methodology. The company notes that base models are improving so fast that reinforcement learning (RL) recipes must adapt constantly. At their current scale of about 3,000 GPUs, one training run produces roughly 33 billion tokens a day, of which about 16 billion are trainable completion tokens after filtering and masking. Mistral’s composable RL library allows a single training run to combine diverse tasks ranging from single-turn chat to complex scientific problem solving and long-horizon tool use.

Mistral’s post notes that the model’s reinforcement learning run is still in flight, and the company expects large and rapid improvements in the coming weeks as training continues on their expanded infrastructure. ML4 is the first milestone funded by their €3 billion Series D equity round. They also indicated that ML4 will serve as a foundation for future specialized and optimized models tailored for critical industries.

For teams deploying autonomous systems, the upcoming open-weight release of this competitive 1T-parameter model will expand the options for building robust, self-hosted environments. You can review how this fits into the broader landscape of Open-Weight Models That Can Actually Drive a Harness. With the impending release of the weights, organizations will be able to evaluate ML4 directly on private cloud or on-premise infrastructure, a critical requirement for those handling sensitive data, strict compliance regimes, or advanced cybersecurity operations where provider control is unacceptable.