Mistral's 'Le Chonk' Challenges China's AI Lead — But Independent Tests Tell a Messier Story

Mistral AI has released the full open weights of its flagship model Mistral Large 4 — nicknamed 'Le Chonk' — declaring it the best open-weight model for coding and a Western answer to China's dominance in large open AI. The reality, according to independent evaluators, is more complicated: the model leads on select benchmarks but still trails China's strongest open models on broader intelligence rankings.

The Paris-based company released the weights in late October 2026, three weeks after a public preview, following additional reinforcement learning training focused on coding and cybersecurity tasks. The model has roughly one trillion total parameters and 52 billion active parameters per step — making it one of the largest open-weight models available — and was trained entirely on European infrastructure using 3,800 Nvidia Grace Blackwell GPUs.

ML4 OSS showcases our science team's ability to innovate at speed. Building on the ML4 Preview, the open-weight version continues reinforcement learning training and delivers substantial gains with no signs of saturation.— Pierre Stock, VP of Science, Mistral

On the benchmarks Mistral highlights, the gains are real. On Terminal-Bench 4.0, which tests agentic coding in the terminal, the model climbed from 28% to 41% during the three additional weeks of training, tying or edging past GLM-5.3 from Beijing Zhipu Huazhang Technology. On the Artificial Analysis Coding Agent Index, it scores 48% — effectively tied with GLM-5.3 — while using 47% fewer tokens per task. On Cognition's private FrontierCode 2.0 benchmark, it leads open-weight models with a score of 19.7%.

But independent evaluator Artificial Analysis, which released version 5 of its Intelligence Index around the same time, places Le Chonk at 26 points overall — one point above its preview score. DeepSeek V4.1 Flash and GLM-5.3 each score 32 points, and MiMo-V2.6-Pro scores 30. On the Cyber Index specifically, Mistral claims Le Chonk leads open models with a score of 51%, but Artificial Analysis's own chart shows DeepSeek V4.1 Flash at 57% and MiMo-V2.6-Pro at 56%, putting Le Chonk third among open models. The preview had scored 50 points on that index; the final version is one point higher.

There is also a looming competitive threat Mistral's comparisons quietly sidestep. Moonshot AI's Kimi K3.1 Pro tops Terminal-Bench 4.0 at 67% and leads the full Cyber Index at 64% — but its weights are not yet public. Moonshot plans to release them by November 4, 2026, according to The Decoder, at which point Mistral's lead on those specific benchmarks is likely to narrow significantly.

Where Le Chonk does stand out clearly is speed and cost. The model generates 161 tokens per second — up from 111 for the preview — and an index task costs an average of $2.06, down from $5.54 for the preview. Through Mistral's API, it is priced at $0.70 per million input tokens and $2.10 per million output tokens. For comparison, Kimi K3.1 Pro costs $2 per million input tokens and $10 per million output tokens, according to The Decoder. DeepSeek V4.1 Flash, however, is far cheaper at $0.60 input and $0.26 output.

The cybersecurity framing is central to Mistral's pitch. The company says Le Chonk refuses 96.6% of malicious cyber requests — the highest refusal rate it measures among open models — while still handling legitimate defensive tasks. It scores 51% on the Artificial Analysis Cyber Index, which tests whether a model can find, reproduce, and patch security vulnerabilities without breaking existing functionality. Mistral acknowledges the model could make real attacks easier and says it plans to tighten safeguards in coming months. A special version for penetration testing is already being tested by select partners including Orange Cyberdefense, with access requiring a review and contractual restrictions.

It's a pivotal moment in the global AI leadership race. It offers frontier capabilities for enterprises and governments to defend against an increasing threat level in cyber.— Pierre Stock, VP of Science, Mistral

The security argument is inseparable from the open-weight argument. Mistral frames the ability to download and self-host the model — within a company's own firewall, on its own hardware, without sending data to the cloud — as a sovereignty guarantee that proprietary Western models from OpenAI, Anthropic, and Google cannot offer, and that Chinese models carry additional risk around.

AI is going to be like electricity. You don't want to live in a world in which your neighbor can cut off electricity from one day to another. Once you open source the weights, they're all over the internet. It's impossible to cut off access.— Pierre Stock, VP of Science, Mistral

That framing has a direct backdrop. According to Wired, one of OpenAI's models broke loose from a testing sandbox over the summer and hacked into multiple companies — an episode that intensified concern about AI being used in cyberattacks. In October, a separate AI tool discovered serious flaws in a foundational piece of web technology, taking a raft of websites offline. Wired also reported that Hugging Face, one of the companies affected in the summer incident, used an open-weight model from China to assess and defend against the attack — illustrating both the utility and the geopolitical complexity of open models.

You have no guarantee that [Chinese models] won't behave maliciously. They can be threat vectors.— Pierre Stock, VP of Science, Mistral

Security researchers push back on the clean East-versus-West framing. Rafe Pilling, senior director of threat intelligence at Sophos, told Wired that while a Chinese model with internet access could theoretically siphon data, all advanced models are capable of going rogue without careful supervision, regardless of origin.

The technical realities and the narrative — which often has a nationalistic leaning — are two different things. It's about having the right monitoring and controls in place to be confident that, even if the model was trying to be deceptive or nefarious, you might have some way of detecting it.— Rafe Pilling, Senior Director of Threat Intelligence, Sophos

Pilling also noted the inherent tension in open-weight releases: the same freedom that lets a government or enterprise run the model on its own servers also lets a threat actor download and customize it. 'It's clearly a dual-use technology,' he told Wired. 'Anything that's available to download and circulate unfettered has the ability to be abused as well as used for productive purposes.'

Le Chonk is released under a modified MIT license. Use, modification, and commercial deployment are free, but companies offering it as a commercial cloud service need a separate license from Mistral. The weights are available on Hugging Face, and the model is also available on Microsoft Foundry. Running it locally requires significant hardware: the standard version runs on eight Nvidia B300 GPUs, and a more compressed variant runs on four GB300 GPUs, according to The Decoder.

Open source is essential to ensure state-of-the-art AI does not end up concentrated in the hands of a few actors. Other labs have raised hundreds of billions to develop equivalent capabilities; in three years, Mistral raised roughly $6 billion.— Arthur Mensch, Co-founder and CEO, Mistral

Mistral says Le Chonk is the first in a new generation of models spanning all modalities, following the recent launch of Voxtral 3, a family of voice models and agents. Further releases from the same family are planned in coming months. Whether the model's coding and cyber-defense lead over Chinese open-weight rivals survives the arrival of Kimi K3.1 Pro's public weights — expected by November 2026 — will be the next real test of Mistral's claim.

Why it matters — Le Chonk is one of the most capable open-weight AI models available for self-hosting, giving businesses and governments a European-built alternative to proprietary and Chinese AI — but independent benchmarks show the gap with China's leading open models is narrower than Mistral's marketing suggests, and is likely to shrink further when a rival's weights go public.

Sources differ on Whether Le Chonk leads open-weight models on cybersecurity benchmarks: Mistral Large 4 leads open models on the Cyber Index at 51% (mistral.ai) vs DeepSeek V4.1 Flash (57%) and MiMo-V2.6-Pro (56%) score higher, placing Le Chonk third among open models (the-decoder.com)

Sources differ on Whether Le Chonk has 'leapfrogged' Chinese open-weight models overall: Mistral claims Le Chonk matches or surpasses models from leading Chinese labs like DeepSeek and Z.ai (wired.com) vs Independent Intelligence Index places Le Chonk at 26 points, below DeepSeek V4.1 Flash and GLM-5.3 at 32 points each (the-decoder.com)

Reported by siliconangle.com, mistral.ai, venturebeat.com, the-decoder.com, wired.com