> ## Content Index > Fetch the complete content index at: https://www.tao.media/llms.txt > Use this file to discover other available public pages before exploring further. # Mistral Releases Large 4 ‘Le Chonk’ Weights With Coding and Cyber Defense Gains - URL: https://www.tao.media/mistral-releases-large-4-le-chonk-weights-with-coding-and-cyber-defense-gains/ - Published: 2026-10-30T13:17:18.000Z - Updated: 2026-10-30T13:17:19.000Z - Description: The trillion-parameter model gives developers and security teams a self-hosted option with more efficient reasoning and enterprise customization. - Author: Bart Hillerich - Tags: Mistral, AI, News Mistral AI has [released the weights of an upgraded Mistral Large 4](https://x.com/MistralAI/status/2115811242556137820?ref=tao.media), its trillion-parameter multimodal model nicknamed "Le Chonk," following three additional weeks of reinforcement learning after its public preview. The release improves coding, defensive cybersecurity, and reasoning efficiency, while allowing organizations to deploy and customize the model on their own infrastructure. The gains combine stronger task completion with shorter outputs. Mistral reports that Large 4 matches the leading open-weight model on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task. On Terminal-Bench 4.0, its score rose by 13 percentage points from the preview, with output-token use falling 30%. The [upgraded model](https://mistral.ai/news/mistral-large-4-oss/?ref=tao.media) is available through hosted services as well as downloadable weights. Developers building coding agents and security teams working with sensitive codebases can choose between managed inference and running the model inside their own environments. ## A Trillion-Parameter Model With Selective Activation Mistral Large 4 uses a mixture-of-experts architecture, which routes each token through a subset of the model rather than activating all its parameters. It has 128 routed experts, four of which are active per token, plus one shared expert. About 49 billion parameters are active per token, or 52 billion when embeddings and output layers are included. The 1T-A52B designation captures this split. The total parameter count describes the full collection of weights, while the smaller active count reflects the portion used for each token. Selective activation reduces computation relative to activating the entire trillion-parameter network, although deployment still needs to accommodate the model's weights. The [Hugging Face model card](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B?ref=tao.media) lists a 1-million-token context window, text and image inputs, and text output. Its vision encoder handles variable image sizes and aspect ratios, supporting work with charts, technical drawings, documents, and other visual material. Mistral says the model is natively fluent in more than 160 languages, including every official European Union language. Mistral trained Large 4 from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its own European data centers. Instruction following, reasoning, coding, and agentic workflows share a single set of weights, with native function calling, JSON output, and strict JSON schema enforcement for applications that need structured responses or external tools. Developers can choose low, medium, or high reasoning effort on each request. Low replaces the previous "None" mode and is intended for simple tasks; medium targets everyday work; and high, the default, is intended for harder tasks requiring more extensive reasoning. ## Coding Gains Combine Quality and Token Efficiency Mistral's reported coding results cover several different measures: whether an agent can complete a task, whether its changes are suitable for merging, and which output human reviewers prefer. | Evaluation | Reported Result | What It Measures or Compares | | -------------------------------------- | -------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------- | | Cognition FrontierCode 2.0 | 19.7% | End-to-end code quality and whether a maintainer would merge an agent's change; Mistral reports a lead among open models. | | Artificial Analysis Coding Agent Index | 48% | Effectively tied with GLM-5.3, with 47% fewer tokens per task on average. | | Terminal-Bench 4.0 | 41% | A 13-percentage-point gain over the preview, with 30% fewer output tokens. | | Surge Code | Preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash | Blind pairwise human evaluations; Claude Opus 5.5 ranked above Large 4 among the models evaluated. | FrontierCode 2.0 is a private benchmark from Cognition. Its focus on whether a maintainer would accept a change tests an agent's work as a repository contribution, rather than only whether it produces plausible code. Token efficiency matters in workflows where an agent repeatedly reads files, reasons through a problem, calls tools, and revises its work. The number of tokens consumed across those steps contributes to inference cost. Mistral's results show improvements in both completion scores and token use, rather than a performance gain achieved solely through longer reasoning outputs. The additional training also improved broader enterprise tasks. Large 4 scored 73.3% on FinWorkBench, or Finch, which includes finance and accounting workflows such as creating and editing spreadsheets. It reached 84.5% on BrowseComp, compared with 75% for an earlier supervised fine-tuning checkpoint. ## Cyber Defense, Controlled Testing, and Customization Large 4 scored 51% on the Artificial Analysis Cyber Index, which tests whether a model can discover vulnerabilities in a codebase, reproduce and validate them, and patch them without breaking existing functionality. Mistral has worked with security organizations evaluating the model in operational settings, including Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik, and Zscaler. Their published comments describe uses ranging from source-code analysis and vulnerability remediation to defensive-log analysis and security investigations. ![Mistral Large 4 customers and partners](https://storage.ghost.io/c/78/0b/780ba906-b1a7-4bf0-873c-bdd5c32e5331/content/images/2026/10/partners-ml4-oss_17QG5o.webp) [Mistral Large 4 customers and partners | Mistral AI](https://mistral.ai/news/mistral-large-4-oss/?ref=tao.media) Downloadable weights let organizations handling proprietary software or regulated data run security analysis on infrastructure they control. Teams can also customize the model around their security data, policies, detection logic, and threat environments through [Mistral Forge](https://mistral.ai/products/forge/?ref=tao.media). Mistral reports an average malicious-cyber-request refusal rate of 96.6%, measured on cyber-related prompts from JailbreakBench, StrongREJECT, and AgentHarm. Separately, Large 4 resisted 97.1% of attacks on the B3 Agent Security Benchmark, up from 93.3% in the preview. That evaluation addresses threats to deployed agents, including system-prompt exfiltration and malicious code injection. The company is adjusting its guardrails after monitoring cybersecurity use during the preview. While it observed refusals of black-box attacks against live targets, Mistral said it also identified potential for the model to help users conduct real attacks. A specialized cybersecurity variant is in preview with vetted enterprise design partners, including Orange Cyberdefense. It targets penetration testing, vulnerability assessment, and red-teaming, with contractual safeguards restricting testing to owned or explicitly authorized systems in controlled environments. ## Deployment Options, License Terms, and API Pricing Large 4 is available in Mistral Studio, through the Mistral API and OpenRouter, and through Microsoft Foundry and NVIDIA-hosted endpoints. Self-hosting options include vLLM and SGLang, alongside an optimized NVIDIA NIM container. Mistral also offers an NVFP4-quantized checkpoint aimed at Blackwell GPUs for a smaller memory footprint. The weights use the [Mistral Ecosystem Open-Weight License](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE?ref=tao.media), a modified MIT license allowing use, modification, fine-tuning, and commercialization. Businesses operating a Model-as-a-Service offering, including through an affiliate, must obtain a separate commercial license before commercial use of the model or its derivatives. The license exempts internal use and access through Mistral's official products from that requirement, and distinguishes general model access from capabilities embedded within specific end-user product features. Mistral's API rates are $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens. The company plans to use Large 4 as the foundation for further specialized and optimized models. That work follows releases including [Voxtral 3's multilingual speech models and voice agent SDK](https://tao.media/mistral-releases-voxtral-3-with-multilingual-speech-models-and-livekit-voice-agent-sdk/?ref=tao.media). Developers can already try Large 4 on repository tasks in Vibe Code, supplying a feature to build or a bug to fix.