Mistral’s Le Chonk is now a top open model for coding, ranking alongside Codex GPT-6.1 Terra

Mistral has released the weights for its Mistral Large 4 model, nicknamed "Le Chonk," following the model's preview launch on Oct. 6. The French AI lab says the open release sets a new bar for Western open models and cements the model's position among the top open options for coding work. The nickname reflects the model's trillion-parameter size.
The headline result is a jump on agentic coding benchmarks. On Terminal-Bench 4.0, ML4's score climbed from 28% in the preview version to 41% at release, an improvement achieved after three weeks of testing with a select group of partners. That 41% edges out Opencode's GLM-5.3 by a single point. On the composite Artificial Analysis Coding Agent Index, which aggregates three relevant benchmarks, ML4 scores 48, one point behind GLM-5.3's 49. Codex GPT-6.1 Terra also sits at 48 overall. According to Artificial Analysis' ranking, ML4 is at the top of its Coding Agent Index among open models globally.
Terminal-Bench is an agentic benchmark, meaning it tests a model's ability to complete multi-step tasks rather than answer single questions, so gains there tend to reflect real improvements in tool use, planning and recovery from errors. Mistral also emphasizes efficiency: the company says ML4 rivals top open models such as GLM-5.3 while using 47% fewer tokens per task on average. That metric matters for teams paying per token, since fewer tokens per completed task translates fairly directly into lower inference bills. A likely trade-off is that token efficiency and raw ceiling performance do not always move together, but the source data suggests Mistral is claiming competitive accuracy at lower cost.
ML4 now ships with three reasoning modes: low, medium and high. This pattern, already common among larger models, lets users trade latency and cost against depth of reasoning depending on task complexity. For product teams, that means a single model deployment can serve both quick, cheap calls and harder agentic sessions.
The release also leans hard into cybersecurity. When Mistral shipped the preview in early October, it advertised cyber defense capabilities along with customizable control and data sovereignty. Over recent weeks the lab worked with developers and leaders to assess those capabilities, and initial testing partners were given a less guardrailed version of ML4 with expanded cybersecurity capabilities, similar in approach to Anthropic's Project Glasswing and OpenAI's rollout of Astra. On CyberGym, a benchmark that measures how well models find memory-safety vulnerabilities in C/C++ open-source projects, reproduce them with proof-of-concept inputs and patch them without breaking tests, Le Chonk scored just above DeepSeek V4.1 Flash.
Mistral co-founder Guillaume Lample said during the Oct. 6 preview briefing that the cyber defense capabilities would let enterprises and governments defend against threat actors who jailbreak closed models to carry out attacks. Mistral frames the core advantage as control: unlike proprietary models, whose access a frontier lab can technically rescind, customers running the open weights keep access under their own control. The company is also among those that signed Nvidia's Open Secure AI Alliance after the Hugging Face breach, a partnership arguing open models are key to democratizing defenses. In practice, the debate over whether open weights are safer or riskier than closed models is unlikely to settle soon; Mistral's argument here is about sovereignty and permanence of access rather than inherently lower risk.
On training, Mistral says ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, deployed in the company's own data centers in Europe. For scale, Nvidia CEO Jensen Huang said on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs. That gap is the subtext of the release: a leaner lab is claiming results near the frontier of open models at a fraction of the compute, and the token-efficiency figures support a lower-cost positioning. Benchmark scores should still be read cautiously, particularly given the incentive to tune for them, but the preview-to-release improvement in a few weeks is a meaningful signal.
Why it matters: Developers and product teams choosing a coding model now have an open-weight option whose benchmark scores sit within a point of GLM-5.3 and Codex GPT-6.1 Terra on a composite index, with a claimed 47% reduction in tokens per task. Self-hosting the weights gives teams data-sovereignty and access guarantees that API-only frontier models do not offer, at the cost of running and maintaining their own infrastructure, which is an inference rather than a stated Mistral claim. The three reasoning modes also make it easier to tune cost against task difficulty within one model.
Based on reporting from the original publisher. Visit the source for full context and later updates.
Publisher excerpt
Mistral has released the weights for Large 4, along with some impressive Artificial Analysis scores.