Ranking

Mistral Large 4 Goes Open Weight, Lands Ahead of Meta but Behind China

co-produced by AI29. October 2026, 21:47

Mistral CEO Arthur Mensch speaks on stage in front of a large audience at the AI Now Summit — Mistral CEO Arthur Mensch at the company's AI Now Summit. © Mistral AI

Europe’s most powerful A.I. model is now available to everyone, yet it still cannot catch China’s open-weight leaders. Mistral AI has just released the weights of Mistral Large 4, nicknamed “Le Chonk.” The model, with roughly one trillion parameters, can now be downloaded from Hugging Face, run on one’s own hardware and further trained. In Artificial Analysis’s redesigned ranking of open models, it lands in eighth place, ahead of Meta but behind seven models from China.

What Mistral Released

Until now, Mistral Large 4 was available only as a preview through Mistral’s application programming interface (A.P.I.). According to Mistral, the model has been trained further since then. The key details:

Size: Roughly one trillion parameters, 52 billion of which are active per request (mixture of experts)
Capabilities: Processes text and images, combines standard answers, reasoning and agentic tasks in a single model, and handles more than 160 languages
Training: Trained from scratch on European infrastructure, according to Mistral on its own cluster in France
License: A modified MIT license allows use, modification and commercial deployment. Companies that offer the model as a hosted service need a separate license from Mistral.
Price: Through the A.P.I., the model now costs about €0.63 per million input tokens and about €1.89 per million output tokens, roughly half the price of the preview.

Mistral is positioning Large 4 primarily for coding agents and cybersecurity, such as finding and fixing vulnerabilities in software. A specialized variant for penetration testing is also available, for now only to selected partners such as Orange Cyberdefense.

Eighth Place Behind China’s Open-Weight Leaders

The more interesting question is where the model ranks. Artificial Analysis recently introduced version 5 of its Intelligence Index, which relies heavily on secret test questions and rates open-weight models far lower than before. At the same time, the benchmarking service redesigned its ranking of open models: It can now be filtered by model size, features a new Openness Index that rates how open a model really is, and tracks the progress of open and closed models over time.

According to Artificial Analysis, the released version of Mistral Large 4 scores 26 points, one more than the preview. This is what the top of the open-model ranking looks like:

DeepSeek V4.1 Flash (China): 32 points
GLM-5.3 from Z.ai (China): 32 points
MiMo-V2.6-Pro from Xiaomi (China): 30 points
Ling 3.1 Flash from InclusionAI, MiniMax-M3.1-Flash and Step 5 Preview from StepFun (all China): 29 points each
MiMo-V2.6-Flash from Xiaomi (China): 27 points
Mistral Large 4 (France): 26 points
Muse Spark Open from Meta (U.S.): 25 points
Qwen3.8 27B from Alibaba (China): 20 points

That makes Mistral Large 4 the best open model from outside China. Meta trails by one point, while U.S. models such as Reflection AI’s Beam 1.0 (18), Nvidia’s Nemotron 3.5 Super (14) and Thinking Machines’ Inkling (12) are well behind. Google’s open model Gemma 4, at 31 billion parameters, is much smaller and has not yet been scored on the new index. Moonshot AI’s Kimi K3.1 Pro is even stronger at 46 points, but its weights have not yet been released.

Its position has not changed since the first test of the preview: Back then, too, Mistral ranked eighth behind seven Chinese models. The gap to the leading closed models is wide. Anthropic’s Claude Opus 5.5 tops the index with 58 points, and OpenAI’s GPT-6 Astra scores 57.

Stronger in Coding and Cybersecurity

Mistral does better in specialized disciplines. On Artificial Analysis’s Coding Agent Index, Large 4 ranks second among open models, one point behind GLM-5.3, while using 47 percent fewer tokens per task, according to Mistral. On the Cyber Index, it scores 51 points, putting it third among open models. Costs have also fallen: According to Artificial Analysis, running one Intelligence Index task now costs about €1.85, compared with about €4.98 for the preview.

Pierre Stock, Mistral’s vice president of science, told Fortune that Large 4 is the best open model for cybersecurity and coding. He said customers in regulated industries mainly want assurance that their access to a model cannot be cut off, which becomes impossible once the weights are public.

How Open and How European Is It?

The picture is not entirely clear. Individual test results vary considerably depending on who measures them: On the agentic benchmark Terminal-Bench 4.0, Mistral reports 41 percent, while Artificial Analysis measured 25 percent. And while the license is more permissive than many rivals’, it is not entirely free, because providers that resell the model as a service need a commercial license from Mistral.

Then there is the question of whether Europe needs its own frontier models at all. Mistral’s chief executive, Arthur Mensch, recently argued that open models from the United States or China will not save Europe, since dependence on other people’s open models is still dependence. Not everyone agrees. Ursula von der Leyen, president of the European Commission, has said Europe can benefit greatly from A.I. without developing frontier technology itself. And the search engine Ecosia dropped Mistral in favor of other open models, including Chinese ones, citing quality and cost.