Mistral has officially released the weights for Mistral Large 4, its most extensive model to date, under the nickname "Le Chonk." The model features approximately one trillion parameters in total, with 52 billion active per step. Trained entirely in Europe, it accepts text and image inputs to generate text and code. The release comes three weeks after a public preview, with the weights now available on Hugging Face under a modified MIT license that permits free use and commercial deployment, though offering the model as a cloud service requires a separate agreement from Mistral.
In terms of performance, Mistral claims Large 4 outperforms its open rivals in coding tasks. Independent testing by Artificial Analysis shows the model scoring 26 points on the Intelligence Index version 5. This places it ahead of the median for similar-sized open models but significantly behind leading Chinese systems. Specifically, DeepSeek V4.1 Flash and GLM-5.3 both score 32 points, while MiMo-V2.6-Pro achieves 30 points. The gap widens further against proprietary leaders like Google's Gemini 4 Carbon (59 points) and Anthropic's Claude Opus 5.5 (58 points).
Despite the overall ranking deficit, Large 4 demonstrates specific strengths in coding benchmarks. On Terminal-Bench 4.0, the model improved from 28 percent to 41 percent after additional reinforcement learning training, narrowly edging out GLM-5.3 at 40 percent. On the FrontierCode 2.0 benchmark, it leads open models with a score of 19.7 percent, though Claude Opus 5.5 scores 46 percent. Blind comparisons conducted by Surge indicate that testers preferred Large 4's code over outputs from GLM-5.3, Kimi K3, and DeepSeek in most matchups, although it only won 22 percent of comparisons against Opus 5.5.
The model also shows notable efficiency and speed improvements. Large 4 generates tokens at a rate of 161 per second, up from 111 during the preview phase. It is faster than most competitors, with only DeepSeek V4.1 Flash and Muse Wonder surpassing it in speed. Cost-wise, Mistral has reduced API pricing to $0.70 per million input tokens and $2.10 per million output tokens, making it cheaper than GLM-5.3 and Kimi K3.1 Pro, though more expensive than DeepSeek V4.1 Flash and MiMo-V2.6-Pro. Additionally, the model is more token-efficient, generating 110 million output tokens across full index tests compared to 150 million in the preview.
In the realm of cyber defense, Mistral positions Large 4 as a leader among open models, yet data suggests otherwise. The company states the final version scores 51 percent on Artificial Analysis' Cyber Index, but their own charts show DeepSeek V4.1 Flash (57 percent) and MiMo-V2.6-Pro (56 percent) ranking higher, placing Large 4 in third place among open models. However, it excels in safety refusals, rejecting 96.6 percent of harmful cyber requests on average, a rate higher than any other open model. Mistral acknowledges the potential for misuse and plans to tighten safeguards, with partners like Orange Cyberdefense already testing specialized versions for penetration testing under strict contractual restrictions.
Looking ahead, Mistral researchers Pierre Stock and Corentin Petit described the development as a "two-year bet" that is now yielding results, noting that performance limits are defined by the difficulty of tasks the team can create and verify. The company admits it operates with fewer resources than US and Chinese labs but views Large 4 as a foundation for future specialized models. While Mistral currently leads in coding among open weights, the upcoming release of Moonshot's Kimi K3.1 Pro on November 4, 2026, which scores 67 percent on Terminal-Bench 4.0, is expected to challenge this position.
Sources: The Decoder.