Mistral releases Le Chonk, its 1-trillion-parameter open-weight model

French AI lab Mistral has released the downloadable weights of Mistral Large 4 (ML4), the trillion-parameter model it nicknames “Le Chonk”, after a preview that began on October 6. The weights arrived on October 29, 2026, and the most important claim attached to them is Mistral’s: that the model now leads open-weight models on agentic coding benchmarks.

Le Chonk is Mistral’s largest model to date. It is a multimodal mixture-of-experts model, an architecture that routes each token through only a subset of the network, so only part of the total parameter count does work at any moment. Mistral pitches it as a foundation for “sovereign AI”: a model that organisations can inspect, customise, self-host, or run through controlled managed services.

What Le Chonk is, in numbers

The name refers to the model’s size: about one trillion parameters. Official figures are not perfectly consistent. A reference guide tracking the model notes that Mistral’s release material gives about 1 trillion total and 49 billion active parameters, while the model documentation lists 1.05 trillion total and 52 billion active, a difference Mistral has not explained.

Specification	What the sources state
Architecture	Mixture of experts
Total parameters	About 1T in the release; 1.05T in model documentation
Active parameters	49B in the release; 52B in model documentation
Context window	1M tokens, per model documentation
Input / output	Multimodal input; text output
Reasoning modes	Low, medium and high
API model ID	mistral-large-4
Training compute	4,000 Nvidia Grace Blackwell GPUs over two months, in Mistral’s own European data centres

Mistral says the model was trained from scratch. “ML4 was trained from scratch on 4,000 Nvidia Grace Blackwell GPUs over two months, deployed in Mistral’s own data centers in Europe,” the company said, as quoted by ZDNET. ZDNET sets that against Nvidia CEO Jensen Huang’s statement on X that OpenAI trained GPT-6 Astra on roughly 100,000 GPUs.

What changed between the preview and the weights release

The gap between preview and weights was about three weeks, during which Mistral tested the model with a select group of partners and ran an additional period of reinforcement learning, a training stage in which the model is rewarded for successful outcomes. Three things changed in that window, according to the coverage:

• Reasoning modes. The model now offers three reasoning modes (low, medium and high), letting users trade answer quality against inference cost by task difficulty.
• Agentic coding scores. ZDNET reports that ML4’s score on Terminal-Bench 4.0, a benchmark for coding agents working in a terminal, rose from the preview’s 28% to 41%. A guide that recorded the preview’s vendor-reported numbers lists the preview at 28.3% on the same test.
• Token efficiency. Mistral says ML4 “rivals top open models such as GLM-5.3 on the Artificial Analysis Coding Agent Index while using 47% less tokens per task on average”.

A Medium post from the CodeToDeploy publication frames the release as making the model about 63% cheaper per task, saying the cost of completing benchmark tasks fell by more than 60% in three weeks. The post’s full workings are not visible in the available text, so that figure should be read as the author’s summary rather than a confirmed measurement.

How Le Chonk’s coding claims stack up against rivals

Most benchmark numbers in circulation come from Mistral. The Le Chonk Guide notes that the preview scores were reported in Mistral’s announcement and attributed to Artificial Analysis, an independent evaluator, and that it has not reproduced the tests or checked the evaluator’s records. The figures below should be read the same way.

• Terminal-Bench 4.0: 41%, one point ahead of GLM-5.3, per ZDNET.
• Artificial Analysis Coding Agent Index: 48, one point behind GLM-5.3’s 49 and level with Codex GPT-6.1 Terra’s 48, per ZDNET. The index is a composite across three benchmarks.
• Cognition’s FrontierCode 2.0: 19.7%, per Mistral.
• Human-preference evaluations: Mistral says the model ranked behind only Claude Opus 5.5 among the models tested.

The sources do not fully agree on how to describe the result. ZDNET says the model ranks at the top of the Artificial Analysis Coding Agent Index among open models globally, yet also reports it one point behind GLM-5.3 on the overall index. HyperAI describes the 48% as “roughly matching the leading open-weight model”. There is a further wrinkle: the guide lists a vendor-reported 49.8% on the Coding Agent Index for the preview, while noting the release did not specify that figure. That preview number is higher than the 48 reported for the final weights, and none of the sources explains the difference. ZDNET also cautions that benchmark scores “should be taken with a grain of salt”.

Mistral also reports results outside coding: 73.3% on Finch for finance workflows such as spreadsheet creation and editing, 84.5% on BrowseComp for web browsing, and resistance to 97.1% of attacks on the B3 Agent Security Benchmark.

How builders can access Le Chonk

Le Chonk is available through Mistral Studio, the Mistral API, OpenRouter, Microsoft Foundry, and NVIDIA-hosted endpoints or NIM containers. The open weights can be downloaded from Hugging Face and served with vLLM or SGLang, two widely used open-source inference engines.

Licence. The weights ship under the Mistral Ecosystem Open-Weight license, which HyperAI describes as a modified MIT-style licence. It permits free use, modification, fine-tuning and commercialisation, except that businesses operating Model-as-a-Service offerings need a separate commercial licence. That carve-out matters for anyone planning to resell hosted access to the model. Open weights and open source are not the same thing, as the Le Chonk Guide points out. Other recent large open-weight releases have also come with custom terms, such as Moonshot’s Kimi K3.1 Pro.

Mistral’s own API pricing. Price is part of Mistral’s efficiency pitch, but the reported figures differ. HyperAI reports $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens on Mistral’s API. The Le Chonk Guide, written during the preview, saw $0.68 input and $2.09 output displayed in the model documentation, alongside higher listed rates of $1.36 and $4.18, with no confirmed end date for the lower rates. Builders should check current rates directly before budgeting.

Self-hosting. Running the weights is a data-centre job. Active parameters describe computation per token, not storage. Using the documented 1.05 trillion total, the guide estimates that even raw 4-bit weights come to about 525 GB, before the memory needed to run the model. The guide had not tested any local setup.

Le Chonk’s cybersecurity pitch

Mistral launched the preview with a focus on cyber defence, advertising customisable control and data sovereignty. During testing, its initial partners received a less guardrailed version of ML4 with “expanded cybersecurity capabilities,” an approach ZDNET compares with Anthropic’s Project Glasswing and OpenAI’s Astra rollout.

“The cyber defense capabilities will enable enterprises and governments to defend themselves against threat actors that are jailbreaking closed models to perform cyberattacks.” (Guillaume Lample, Mistral co-founder, at the October 6 preview briefing)

The reported results:

• Le Chonk scored just above DeepSeek V4.1 Flash on CyberGym, a defensive benchmark in which models find real memory-safety bugs in C/C++ open-source projects, reproduce them and patch them without breaking tests.
• Mistral reports 51% on the Artificial Analysis Cyber Index, which covers vulnerability discovery, reproduction and patching.
• Mistral says the model refused malicious cyber requests 96.6% of the time on average across jailbreak, refusal and agent-harm evaluations.

Mistral is offering vetted enterprise partners a specialised cybersecurity variant for penetration testing, vulnerability assessment and red teaming in authorised environments, and customers can adapt the model further with Mistral Forge using their own security data, policies, detection logic and threat environments. Mistral says it will keep tightening guardrails after monitoring live traffic. It also argues that open weights keep security under the user’s control, unlike proprietary models whose access frontier labs can withdraw. Mistral is one of the companies that signed Nvidia’s Open Secure AI Alliance, which argues that open models are crucial to spreading AI security defences.

What Le Chonk means for builders, and the open questions

For teams building coding agents, Le Chonk adds a Western-made option near the top of the open-weight field. Mistral’s headline efficiency claim, 47% fewer tokens per task, matters as much as the scores, because agent workloads burn tokens across many steps. The guide makes the same point from the other side: a cheaper token does not always mean a cheaper finished task, since retries, long answers and failed tool calls change the bill. Recent open coding releases such as Ricursive’s Ripple-65B make the comparison set crowded, so testing on real workloads matters more than any single index.

Open questions the sources leave unresolved:

• Spec discrepancies. The 49B versus 52B active-parameter figures, and 1T versus 1.05T total, remain unexplained.
• Independent verification. The benchmark figures are vendor-reported or attributed by Mistral to Artificial Analysis; none of these sources reproduced them.
• Pricing. Two sets of rates have been displayed, with no confirmed end date for the lower one, and a later report gives slightly different numbers.
• Dual-use risk. A model tuned to find and patch vulnerabilities could also help find them for attack; Mistral’s answer so far is refusal training, ongoing guardrail updates and restricting the stronger variant to vetted partners.

What to watch next

The next signals to look for are independent results on the released weights rather than the preview: third-party runs of the Artificial Analysis indices, and community reports from teams serving the model with vLLM or SGLang. Watch for Mistral clarifying the parameter counts and which API rates apply going forward, for any changes to the guardrails Mistral says it will tighten after monitoring live traffic, and for how the commercial licence for Model-as-a-Service providers shapes which hosts offer Le Chonk.