Mistral Large 4 released with open weights

Summary

Mistral has released the open weights of its 1-trillion-parameter MoE model, Large 4 (Le Chonk), on Hugging Face. Following three weeks of additional RL training, performance on Terminal-Bench 4.0 improved from 28% to 41%. However, the company acknowledges that leading Chinese open models still maintain their lead on the general Artificial Analysis index.

Why it matters

Using only ~49 billion active parameters per token in a 1-trillion-parameter model represents a significant efficiency milestone for the open-weight ecosystem. Additionally, API prices have been halved ($0.70/$2.10 per M tokens), lowering the barrier to entry for developers. In contrast, the new 'Mistral Ecosystem Open-Weight License' includes a commercial exception for service resellers. This makes the license significantly less permissive than Large 3's Apache 2.0 license and complicates the 'open' label.

Highlights

• MoE architecture: 128 experts, with only 4 active per token → 49 billion active parameters out of a 1-trillion total
• Terminal-Bench 4.0: Improved from 28% to 41% after 3 weeks of RL; outperforms DeepSeek V4.1 Flash (31%) and Kimi K3 (21%)
• The new license prevents companies from reselling model access as a service without a separate Mistral agreement

Critical take

Mistral's own documentation acknowledges that Chinese open models (GLM-5.3, Qwen, Kimi) outperform Le Chonk on the general Artificial Analysis index. Ecosia's switch to these Chinese models weeks ago also weakens the narrative of a resurgence for Western open-weight models. The 'open-weight' framework also feels forced, as the license explicitly restricts serving the model via API, which is the most common form of commercial deployment.