Mistral Large 4 Le Chonk: Big Claims, Thin Margins vs China

What Mistral launched

French AI lab Mistral has previewed Mistral Large 4 (ML4), a trillion-parameter open-weight model it has nicknamed "Le Chonk." The company says ML4 ranks among the best open models in the world and is the strongest open-weight model built outside China by a "substantial margin." 5 It is Mistral's first major model release in roughly five months. 2

The rollout is staged. Developers, "cybersecurity leaders" and state authorities get preview access first. 5 Security experts and government bodies will test a version with fewer safety restrictions ahead of a public release set for October 27. 2 According to a comparison of open flagships, ML4 is currently available through a preview API. Weights are expected at month-end, and no license has been announced yet. 3

Cybersecurity as the selling point

Mistral is pitching Le Chonk mainly as a defensive security tool that customers can own outright. Its argument is that closed models from frontier labs can have access revoked at any time, while open weights stay under the user's control without vendor lock-in. 1 Co-founder Guillaume Lample framed the model as a way for enterprises and governments to defend against attackers who jailbreak closed models to run cyberattacks. 1 Mistral also highlights coding, manufacturing, finance and multimodal tasks as strengths. 5

The framing is political as much as technical. Access to open models, where Chinese labs have been the most capable, has become a flashpoint in the US–China technology rivalry. Western developers are trying to close that gap. 5 Mistral is selling ML4 partly as a European sovereignty play. It says the model was trained on 4,000 Nvidia Grace Blackwell GPUs over two months in its own European data centers. 5 Mistral also says that is fewer GPUs than OpenAI used for Astra, though it claims ML4 is still competitive. 1

What the numbers actually show

The central question is whether Le Chonk beats the Chinese open-weight leaders, and the evidence so far is thin. CEO Arthur Mensch told a launch event in Abu Dhabi that the model surpasses Chinese models "on certain aspects, including cyber." He said this disproves the idea that Europe cannot compete, but he did not name the models or benchmarks he meant. 2 Mistral has not published specific cybersecurity benchmarks. 2

The benchmarks Mistral has released are self-reported, and they show narrow margins. 4 On the DeepSWE agentic coding test, ML4 scores 61.7%, against 61% for GLM-5.3 and 57% for DeepSeek V4 Pro. 4 On FinWorkBench it ties DeepSeek V4 Pro at 67%. 4 In a blind coding evaluation run with Surge AI, it scored 3.74 out of 5. That is above GLM-5.3 and Kimi K3 but well behind Claude Opus 5 at 4.22. 4 Its clearest wins are on visual grounding, where it tops GPT-6 Astra, and on a legal-agent benchmark with 15% against Kimi K3's 13%. 4 Even Mistral's own announcement concedes the model trails the frontier in areas such as coding. 5

No independent benchmarks based on the open weights exist yet. Until the weights ship, nobody outside Mistral can confirm its claims. 4

Size and deployment context

On paper, Le Chonk is not the largest model in its class. It has 1.05 trillion total parameters with about 52 billion active per token. 3 By comparison, Kimi K3 has 2.8 trillion total and 104 billion active, and Qwen3.8's open flagship has 2.4 trillion total and 95 billion active. 3 DeepSeek's V4.1 Flash is a 552-billion-parameter model released under the permissive MIT license. 3

Total parameters drive memory requirements, and active parameters drive per-token cost. 3 A smaller active count could make ML4 cheaper to run than the biggest Chinese rivals. Whether that matters to buyers depends on the license terms Mistral eventually announces.

Our read

Le Chonk looks like a credible re-entry for Mistral, not a clear takeover of the open-weight lead. The self-reported results place it roughly level with GLM-5.3 and DeepSeek V4 Pro on coding and finance. Those margins could easily disappear or reverse under independent testing.

The cybersecurity pitch is the more distinctive part of the launch, and also the least supported. The vetted early-access program for defenders is a sensible step, but the main capability claim rests on unpublished numbers.

The real test comes when the weights and license land. If outside evaluations confirm that ML4 matches China's best at a lower serving cost, Europe has a genuine contender. If they don't, Le Chonk will be remembered mostly as a sovereignty argument with competitive but unremarkable benchmarks.