RELEASE · MODELS · #2778
Mistral releases Mistral Large 4 (Le Chonk) weights — 1T-parameter open multimodal model
Mistral released the weights for Mistral Large 4 (Le Chonk), a 1‑trillion‑parameter natively multimodal model, claiming leading open‑weight results on coding and defensive cybersecurity benchmarks and new three-mode reasoning options to trade off cost and performance. The company says the model was trained on European infrastructure, is fully open‑weight, shows gains on FrontierCode (19.7%) and artificial analysis indexes (coding 48%, cyber 51%), and achieves a 96.6% average refusal rate on malicious cyber prompts while supporting self-hosted deployment.
KEY POINTS
- Mistral released the weights for Mistral Large 4 (Le Chonk), a 1‑trillion‑parameter natively multimodal model, claiming leading open‑weight results on coding and defensive cybersecurity benchmarks and new three-mode reasoning options to trade off cost and performance.
- The company says the model was trained on European infrastructure, is fully open‑weight, shows gains on FrontierCode (19.7%) and artificial analysis indexes (coding 48%, cyber 51%), and achieves a 96.6% average refusal rate on malicious cyber prompts while supporting self-hosted deployment.
- An openly released 1T-parameter multimodal model that claims frontier coding and defensive‑security performance matters because it expands high‑capability, self‑hostable models for enterprises and sovereign AI efforts while shifting cost/performance tradeoffs.
DECISION BRIEF
CONFIDENCE
WHAT CHANGED
Mistral publicly released the weights for Mistral Large 4 (’Le Chonk’): a natively multimodal model with ~1 trillion total parameters (49B active parameters in the published Mixture‑of‑Experts variant) and a 1M‑token context window. The release is fully open‑weight, distributed under Mistral’s open‑weight license, and includes multiple variants and community GGUF builds. Mistral also added three reasoning modes aimed at trading off cost and performance and published benchmark claims showing gains on coding and defensive‑security tasks.
WHY NOW
An openly published 1T‑parameter multimodal model with claimed frontier results and self‑hosting options changes the public availability of high‑capability models: it provides a large, deployable model footprint that Mistral positions for coding and defensive security use cases and for sovereign/local deployments. The release is timed with Mistral’s reported reinforcement‑learning improvements and community distribution on platforms like Hugging Face.
WHO IS AFFECTED
Directly affected parties (per sources): developers and organizations building coding agents and defensive‑security tools; security defenders and incident response teams who may prefer self‑hosted/open weights; platforms and communities that host and distribute models (e.g., Hugging Face, llama.cpp/Unsloth builders); and parties tracking open‑weight model leadership in the public AI ecosystem.
CONFIRMED
- Mistral released the weights for Mistral Large 4 (Le Chonk) and states it is a 1‑trillion‑parameter, natively multimodal model (source: Mistral). (3271) - Mistral says the model is fully open‑weight, trained from scratch on European infrastructure, and is available for self‑hosted deployment (3271). - Mistral reports new reinforcement learning improvements and three reasoning modes to trade off cost and performance (3271). - Mistral reports leading open‑weight results on coding benchmarks, including a FrontierCode score of 19.7% and a 48% score on the Artificial Analysis Coding Agent Index (3271). - Hugging Face documentation states the model has ~1T total parameters, ~49B active parameters in MoE variants, a 1M‑token context window, two official variants, and community GGUF builds; it reports 82% on the Artificial Analysis Cyber Index in Mistral’s testing and positions the model for defensive security work (3275). - Media reporting (WIRED) states Mistral claims Le Chonk is the “world’s best” open‑weight model for automated coding and cybersecurity and that the finalized model is freely available (3273).
UNCERTAIN
Missing or conflicting evidence (explicitly noted): - Independent third‑party validation and peer‑reviewed benchmarking of Mistral’s claims are not present in the supplied excerpts; Mistral’s scores are company‑reported and Hugging Face cites Mistral’s testing (3271, 3275). - The broader comparative ranking against closed‑weight frontier models is asserted by Mistral and reported by WIRED, but no independent benchmark or methodology details are provided in the excerpts to confirm that claim (3271, 3273). - Risk trade‑offs (whether open weights materially increase attacker capabilities versus defensive benefits) are discussed as a tension in reporting but are not resolved by the supplied excerpts (3273, 3275). - Exact refusal rates on malicious cyber prompts (a 96.6% figure appears in the event summary but is not present in the supplied source excerpts) are not confirmed by these excerpts; Hugging Face notes the model still declines harmful requests but the full statistic is not shown (3275). - Details on dataset composition, exact training compute, and full benchmark methodologies are not included in the excerpts and therefore remain unverified (3271, 3275).
WHAT TO WATCH
Concrete observable signals to monitor next (can be publicly checked): - Publication of independent benchmark results (FrontierCode, Artificial Analysis indexes) or third‑party evaluations that reproduce or contest Mistral’s scores. - Availability and download/usage metrics on Hugging Face and community GGUF/llama.cpp mirrors; number of forks, downloads, and implementations. - Third‑party security analyses or red‑team reports describing refusal rates, vulnerability‑reproduction behavior, and malicious‑prompt handling. - Announcements or benchmark releases from other labs (closed or open) responding to Mistral’s claims or publishing head‑to‑head comparisons. - Any public incidents, misuse reports, or documented customizations of the weights that alter capabilities or safety behavior. - Updates to license text, terms of service on distributor platforms, or ecosystem tooling for enforcement of Mistral’s model‑as‑a‑service limitations.
WHY IT MATTERS
An openly released 1T-parameter multimodal model that claims frontier coding and defensive‑security performance matters because it expands high‑capability, self‑hostable models for enterprises and sovereign AI efforts while shifting cost/performance tradeoffs.
EVIDENCE MAP
4Editorial claims linked to specific sources, with support, contradiction and context shown separately.
Mistral released the weights for Mistral Large 4 (nicknamed Le Chonk).
SUPPORTEDChecked 2026-10-29 · 3 supporting
Hugging Face blog describes the model (Le Chonk) and points to checkpoints in the Mistral Large 4 collection.
SUPPORTS Mistral AIMEDIA · 2026-10-29Mistral's announcement states they are releasing the weights of Mistral Large 4 and describes the public preview and weight release.
SUPPORTS WIRED AIMEDIA · 2026-10-29WIRED reports that Mistral made the model available and discusses the company's statements about the release.
Mistral Large 4 is described as a ~1‑trillion‑parameter, natively multimodal model.
SUPPORTEDChecked 2026-10-29 · 2 supporting
The model is released as open-weight under Mistral’s open-weight license and checkpoints are publicly available.
SUPPORTEDChecked 2026-10-29 · 2 supporting
Hugging Face notes the model is distributed under the Mistral Ecosystem Open-Weight License and points to the available checkpoints and community GGUF builds.
SUPPORTS Mistral AIMEDIA · 2026-10-29Mistral states the model is “fully open-weight” and positions it as an open release for developers and institutions.
Mistral says the model leads open-weight benchmarks for coding and defensive cybersecurity and reports a 19.7% score on Cognition’s FrontierCode 2.0 benchmark.
SUPPORTEDChecked 2026-10-29 · 2 supporting
Hugging Face highlights the model’s strength for cybersecurity and coding agents in its overview.
SUPPORTS Mistral AIMEDIA · 2026-10-29Mistral describes taking the lead on FrontierCode and reports a 19.7% score on Cognition’s FrontierCode 2.0.
CONTEXT WIRED AIMEDIA · 2026-10-29WIRED summarizes Mistral’s claim that Le Chonk is now the best open-weight model for automated coding and cybersecurity and quotes company representatives on this point.
SOURCES & TIMELINE
5Mistral released Mistral Large 4 (Le Chonk), a 1-trillion-parameter multimodal model, now available with improved coding, cybersecurity, and reasoning efficiency. It leads open-weight benchmarks in coding and defensive security while offering three reasoning modes for cost-performance balance. Fully open-weight and European-trained, it empowers sovereign AI with customizable, self-hosted deployment options. Earlier …
Mistral Large 4 (aka Le Chonk) is a multimodal, multilingual, Mixture-of-Experts (MoE) model from Mistral. It takes text and images as input and generates text. It has about 1T total, 49B active parameter , and a context window of 1M tokens . The model is released in two official variants: There are also GGUFs created by Unsloth and the llama.cpp team. It’s distributed under the Mistral Ecosystem Open-Weight Licens…
A few weeks ago, French company Mistral released an early version of its new artificial intelligence model, nicknamed Le Chonk. At the time, the company said it was the most advanced open-weight offering outside of China. After further refinement, Mistral now says Le Chonk is the “world’s best” open-weight model for automated coding and cybersecurity. The advances, Mistral says, were made possible by feeding the mod…
French AI company Mistral has released the weights for its trillion-parameter open source AI model Mistral Large 4 (ML4), allowing businesses and governments to download and deploy the system themselves. Nicknamed “Le Chonk,” the model was first introduced earlier this month . The release comes as competition between open-weight AI developers like Mistral and closed-model providers like OpenAI and Anthropic heats up…
Mistral has released the weights for Large 4 ("Le Chonk"). The model has about one trillion parameters, 52 billion of them active, and can run on your own hardware under a modified MIT license. In independent benchmarks, the model does well on coding. On Artificial Analysis' Intelligence Index, though, it scores 26 points, behind open Chinese models like DeepSeek V4.1 Flash and GLM-5.3 and far behind the leading sys…