On Oct 29 Mistral released an improved Mistral Large 4 (Le Chonk) with open weights on Hugging Face (plus an NVFP4 quantized checkpoint), served on its API as mistral-large-4. After about three more weeks of reinforcement learning since the early-October preview, Mistral reports 19.7% on FrontierCode 2.0, Terminal-Bench 4.0 up from 28% to 41%, and new low/medium/high reasoning modes. The license is a modified MIT that allows commercial use, but running a model-as-a-service business needs a separate commercial license.
Key Takeaways
- ✓Specs: about 1T total parameters with 49B active per token (about 52B including embeddings and output layers), 128 routed experts with 4 active plus 1 shared expert, 1M context, text and image input
- ✓Per Mistral: 19.7% on Cognition's FrontierCode 2.0, ahead of DeepSeek V4.1 Flash at 19.3% and GLM-5.3 at 17.8%, first among open models
- ✓48% on the Artificial Analysis Coding Agent Index, roughly tied with GLM-5.3 while using 47% fewer tokens per task on average; Terminal-Bench 4.0 rose from 28% in preview to 41%
- ✓51% on the Artificial Analysis Cyber Index, and a 96.6% average refusal rate on malicious cyber requests, which Mistral says is the highest among open models
- ✓Trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters; in Surge Code blind pairwise tests it ranked behind only Claude Opus 5.5
Key Decision Metrics at a Glance
Turn your technical choice into a development budget
Compare 40 dev plans & simulate token costs vs $20/mo subscriptions
Project Links & Resources
Direct AccessIn-Depth Technical Analysis
Background and the problem it targets
Mistral put Mistral Large 4 (Le Chonk) into public preview in early October and promised open weights by month-end. Today it delivered, releasing an improved version trained with about three more weeks of reinforcement learning. For European governments and regulated industries, the issue has been that the strongest coding and security models are closed and send data abroad. Mistral is pitching a flagship model you can download, audit and self-host.
Architecture and how it works
Per the Hugging Face model card: about 1T total parameters with 49B active per token (about 52B including embeddings and output layers), a granular MoE with 128 routed experts (4 active) and 1 shared expert, 1M context, text and image input and text output. The checkpoint ships with an embedded DSpark speculative decoding draft head, and an NVFP4 quantized version targets Blackwell GPUs. Reasoning effort has three levels, low, medium and high (default), with low replacing the old no-reasoning mode.
Benchmarks and measured results
All figures below are Mistral-reported: 19.7% on FrontierCode 2.0, first among open models; 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 while using 47% fewer tokens; Terminal-Bench 4.0 up from 28% to 41% with 30% fewer output tokens; 51% on the Cyber Index and a 96.6% refusal rate on malicious requests. Mistral also says live traffic showed potential uplift for real attacks and it is tightening guardrails.
Getting started for developers
Weights are at mistralai/Mistral-Large-4-1T-A52B on Hugging Face, with vLLM, SGLang, transformers and llama.cpp support; or call mistral-large-4 on the Mistral API. Mistral recommends temperature=1.0 and top_p=0.95, high effort for complex agentic work and medium for everyday use. The license is a modified MIT that allows commercial use, but offering it as a public model API service requires a separate commercial license from Mistral.
Benchmark side-by-side against alternatives, or calculate monthly token cost vs subscription break-even.
Discussion & Comments
0Sign in to join the discussion
Connect with AI developers to exchange benchmark insights.