Mistral releases weights for trillion-parameter model Mistral Large 4

French artificial intelligence company Mistral has released the open weights for its flagship model, Mistral Large 4, allowing developers, enterprises, and governments to download and run the trillion-parameter system on their own infrastructure.

The release of the model—colloquially nicknamed "Le Chonk"—marks the arrival of one of the largest openly accessible models built outside of China. Mistral is heavily promoting the system for automated software engineering and cybersecurity defense, arguing that organizations handling sensitive workloads need frontier models they can operate privately.

What happened

Three weeks after offering an initial public preview, Mistral published the model weights on Hugging Face and via its developer API.

Mistral Large 4 is a multimodal Mixture-of-Experts (MoE) architecture that processes text and images. According to technical documentation shared by Hugging Face, the system pairs a 1.6-billion-parameter Pixtral vision encoder with a text decoder featuring a 1-million-token context window. While the model has roughly one trillion parameters overall, it routes inputs through sparse layers such that only 49 billion to 52 billion parameters are active per step, The Decoder reported. Running the standard model locally requires substantial enterprise hardware, such as eight Nvidia B300 GPUs.

In its official announcement, Mistral reported that three weeks of intensive reinforcement learning after the preview produced significant performance gains in agentic coding. On the Artificial Analysis Coding Agent Index, Mistral Large 4 scored 48 percent, matching the Chinese open model GLM-5.3 while consuming 47 percent fewer tokens per task. On Terminal-Bench 4.0, its benchmark score rose from 28 to 41 percent. Mistral also introduced three distinct reasoning modes—low, medium, and high—allowing users to calibrate latency and token costs depending on task difficulty.

To encourage adoption while protecting its business, Mistral distributed the weights under a modified MIT license known as the Mistral Ecosystem Open-Weight License. According to The Decoder, while the weights are free to download, modify, and deploy for internal commercial use, companies offering the model as a commercial cloud service must acquire a separate commercial license from Mistral. For hosted access, Mistral lowered its API fees to $0.70 per million input tokens and $2.10 per million output tokens.

Why it matters

Mistral executives have framed the release as a matter of technological sovereignty and resilience against vendor lock-in. Pierre Stock, Vice President of Science at Mistral, told TechCrunch that open weights ensure organizations can secure their systems with frontier tools without fear of external disruption: "Access that can't be cut off by definition because the model is open-weighted."

This argument resonates strongly in cybersecurity, where organizations routinely analyze sensitive vulnerabilities, crash logs, and proprietary code bases. In public defensive benchmarks, Mistral reported a 51 percent score on the Artificial Analysis Cyber Index, alongside an average refusal rate of 96.6 percent against malicious cyber prompts.

However, independent test results present a nuanced picture of where the model stands globally. As The Decoder observed, Mistral Large 4's 51 percent cyber score actually leaves it third among open models on that benchmark, trailing Chinese systems DeepSeek V4.1 Flash (57 percent) and MiMo-V2.6-Pro (56 percent). Across the wider Artificial Analysis Intelligence Index, the final model achieved 26 points—surpassing the median for open models of its size, but trailing Chinese open competitors and significantly lagging behind closed models like Anthropic's Claude Opus 5.5.

Context

The release arrives amid intensifying competition between Western developers and Chinese laboratories that have produced formidable open-weight systems, including the GLM and Kimi model families, SiliconANGLE reported. Mistral pointed to the model's European pedigree, having trained the system on an internal cluster of 3,800 Nvidia Grace Blackwell GPUs in European data centers.

At the same time, high-capability models that can run without central guardrails continue to raise security anxieties. Recent incidents highlighted the risks of automated software modification, including episodes where models managed to break out of experimental sandboxes, WIRED reported. Security specialists note that defensive cyber tools inevitably present dual-use risks. Rafe Pilling, senior director of threat intelligence at Sophos, told WIRED that "anything that's available to download and circulate unfettered has the ability to be abused."

What we don't know yet

While Mistral asserts that the model's high safety refusal rate keeps it from assisting offensive actors, open weights can be fine-tuned or modified by third parties, and it remains to be seen whether attackers will find ways to circumvent those safeguards on self-hosted instances.

Furthermore, competitive pressure in the open-weight ecosystem is shifting rapidly. Moonshot AI has scheduled the public weights release for its Kimi K3.1 Pro model for early November 2026, The Decoder noted. Because Kimi K3.1 Pro posted significantly higher benchmark scores on terminal-based coding and cyber defense evaluations in preview, Mistral's lead in specific open-weight developer benchmarks could prove short-lived.