Skip to content
trendoviz

Mistral releases Le Chonk, a 1M-context open-weight model

Mistral has released Mistral Large 4, also known as Le Chonk, on Hugging Face. The multimodal, multilingual mixture-of-experts model accepts text and images, has about 1 trillion total parameters with 49 billion active, and carries restrictions on model-as-a-service use.

26 out of 100Dormant7D: Insufficient history ProvisionalPublished

Overview

Mistral Large 4 is available on Hugging Face in two official variants, with additional GGUF distributions from Unsloth and the llama.cpp team. It accepts text and images as input and produces text, supports a 1 million-token context window, and runs under the Mistral Ecosystem Open-Weight License, an MIT derivative that limits model-as-a-service offerings [1].

The release extends the open-weight model field toward longer-context, multimodal workloads. Mistral positions Le Chonk for coding agents, cybersecurity, and multimodal tasks [1].

Why it matters

The combination of image and text input, multilingual coverage, and a 1 million-token context window gives builders and organizations a downloadable model suited to long documents, codebases, and mixed text-and-image workflows [1].

Open weights also place deployment and usage policy under the adopter's control, which Mistral argues matters during incident response when API access can be interrupted [1]. Researchers can likewise measure refusal behavior directly rather than relying on a provider's description of it [1].

Security teams may find the Cyber Index result relevant: Mistral Large 4 scored 82% on a test requiring the model to reproduce a real vulnerability in open-source software and then patch it [1]. The announcement contains no independent third-party test report to corroborate that figure, so it indicates a capability worth evaluating rather than established production performance.

Opportunities

Builders could use the model for local assistants over large codebases, image-and-text document analysis, and security tooling that reproduces a vulnerability and proposes a patch [1]. Mistral suggests fine-tuning smaller models with TRL from Mistral Large 4 traces for codebase patching, and using reproduce-then-patch tasks as reinforcement-learning environments because they offer a verifiable reward [1].

For organizations evaluating refusal policy, the open-weight format allows direct measurement on cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm rather than accepting a provider's characterization [1]. The published coding results also give potential adopters a starting point for comparisons: 61.7% on DeepSWE v1.1, 59.4% on SWE-Atlas-QnA, 28.3% on Terminal-Bench 4, a combined Coding Agent Index of 49.8%, 59.9% on AutomationBench, and 1,393 Elo on AA-Briefcase [1].

Risks

Most published performance data comes from Mistral's own announcement. There is no independent third-party test report in the available material confirming the cyber, coding, or agentic scores, and the Artificial Analysis Cyber Index methodology is not described [1]. Multimodal capability is documented architecturally, but no image-understanding performance metrics are provided [1].

Refusal behavior limits some offensive or dual-use workflows. Mistral reports that Le Chonk's average refusal rate on cyber prompts from JailbreakBench, StrongREJECT, and AgentHarm is higher than that of any other open-weight model [1].

The license is not equivalent to unrestricted open source. The Mistral Ecosystem Open-Weight License is an MIT derivative with restrictions on model-as-a-service offerings, and its commercial terms for many use cases are not provided [1].

Self-hosting shifts hardware, serving, and maintenance work to the adopter. Runtime requirements and practical costs at the 1 million-token context length are unknown, and the roughly 1 trillion-parameter total size is likely to make inference and fine-tuning resource-intensive, though no specific figures are available [1].

Context

Le Chonk uses sparse computation at two levels. In the language model, a router selects a small group of feed-forward experts for each token; separately, a sparse attention indexer chooses which earlier token positions the model attends to [1].

The text decoder has 72 layers, a hidden dimension of 6,144, and 64 attention heads, combining Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that selects up to 2,048 key positions per query [1]. Three decoder layers use RoPE and every fourth uses NoPE, with KV state shared across layers. Beyond the first layer, 71 layers use mixture-of-experts blocks: a softmax router picks four of 128 experts, and one shared expert always runs alongside them, for five active experts per token [1].

The model uses LeRope, a YaRN-scaled rotary position embedding with a learned multiplier per frequency [1]. Its vision side is a 1.6 billion-parameter Pixtral encoder with 30 layers, a hidden dimension of 2,048, and 16 attention heads, operating on 14 × 14 patches with a spatial merge size of 2 [1].

Reported inference support includes Transformers, vLLM, llama.cpp, TRL fine-tuning, inference endpoints, and Dell Enterprise Hub [1].

Known, claimed, inferred

As of October 29, 2026, Mistral Large 4 ranked fifth among open-weight models on the Artificial Analysis Intelligence index, and its combined Coding Agent Index score of 49.8% placed it ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max [1].

Vendor claim: Mistral describes the model as particularly strong in cybersecurity, coding agents, and multimodality [1].

Vendor claim: Mistral reports that several closed frontier models score near zero on the same Cyber Index test because they decline to attempt the work, and that internal testing found the model useful for malware analysis, vulnerability prioritization, and writing detection rules despite no explicit training for those tasks [1].

Hypothesis: The release broadens the set of self-hosted multimodal and coding options, but the announcement alone does not establish cost, latency, or production-reliability advantages over alternative open-weight models [1]. The two forms of sparsity, expert routing and sparse attention selection, point to a design aimed at handling long contexts while limiting per-token computation [1].

Limitations & uncertainties

No independent third-party test results are available to corroborate the reported cyber, coding, or agentic scores, and the Artificial Analysis Cyber Index methodology is not described [1]. Multimodal capability is documented architecturally, but no image-understanding performance metrics are provided [1].

The exact release date, repository identifiers for the two official variants, pricing, licensing cost, and commercial terms are missing [1]. Hardware and runtime requirements, including the practical cost of serving a 1 million-token context, are also unknown [1]. No user or developer testimonials establish how the model performs in real deployments, and details of Mistral's internal security testing are not provided beyond the stated use cases [1]. The source is an official Mistral announcement on Hugging Face, so its capability statements and comparative claims should be distinguished from independently verified findings [1].

Evidence & sources

The key statements of this analysis, labeled by type, with the sources they rely on. Trendoviz does not run product tests.

  • Primary-source or vendor statementMistral Large 4, also called Le Chonk, is a multimodal, multilingual mixture-of-experts model on Hugging Face that accepts text and images and generates text.[1]
  • Primary-source or vendor statementThe model has about 1 trillion total parameters, 49 billion active parameters, and a 1 million-token context window.[1]
  • Primary-source or vendor statementIt is released in two official variants, with additional GGUF formats from Unsloth and the llama.cpp team.[1]
  • Primary-source or vendor statementThe model is distributed under the Mistral Ecosystem Open-Weight License, an MIT derivative that limits model-as-a-service offerings.[1]
  • Primary-source or vendor statementMistral Large 4 scored 82% on the Artificial Analysis Cyber Index, a test requiring the model to reproduce a real vulnerability in open-source software and patch it.[1]
  • Primary-source or vendor statementMistral claims the model is especially strong for cybersecurity, coding agents, and multimodality, and reports that internal testing found it useful for malware analysis, vulnerability prioritization, and detection-rule writing.[1]
  • Primary-source or vendor statementThe combined Coding Agent Index score is 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, with 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4.[1]
  • HypothesisThe release broadens self-hosted multimodal and coding options, but the announcement does not establish cost, latency, or production-reliability advantages over alternative open-weight models.[1]

Sources

  1. 1
    Welcome Le Chonk: Mistral Large 4 on Hugging Face
    ●huggingface.coretrieved Nov 18, 2026

Updates & corrections

  1. First published · published automatically after passing every check

Found an error? Report it.