Mistral AI has released the downloadable weights for Mistral Large 4, its trillion-parameter flagship model nicknamed “Le Chonk,” giving developers, enterprises, and public institutions the ability to inspect, fine-tune, and self-host one of the most capable open-weight AI systems announced to date. The release arrived on October 29, 2026, roughly three weeks after the model entered public preview.mistral
The move matters because it shifts Mistral Large 4 from an API-only preview into a model that organizations can deploy inside their own infrastructure—an option increasingly important for coding agents, cybersecurity workflows, and data-sensitive enterprise applications.mistral
What Mistral released
Mistral Large 4 is a natively multimodal, multilingual Mixture-of-Experts (MoE) model. According to Mistral’s documentation, it has 1.05 trillion total parameters, activates 52 billion parameters during inference, includes a 1.6 billion-parameter vision encoder, and supports a context window of up to 1 million tokens.docs.mistral
The weights are available on Hugging Face in FP8 and NVFP4 variants, with community GGUF builds also available. Mistral says day-zero support is provided through Transformers, vLLM, and llama.cpp.huggingface
Developers can access the model in several ways:
Download the weights from Hugging Face for self-hosting.
Use Mistral Studio, the Mistral API, or OpenRouter.
Deploy through NVIDIA-hosted endpoints or an optimized NVIDIA NIM container.
Customize the model with Mistral Forge or run it through Microsoft Foundry.mistral
Why “open weight” is not the same as open source
Mistral Large 4 is open-weight, not fully open source in the conventional software sense. The model is distributed under the Mistral Ecosystem Open-Weight License, described by Mistral as a modified MIT license.mistral
Under that license, users may use, modify, fine-tune, and commercialize the model. However, companies that operate a Model-as-a-Service offering—reselling hosted inference—must obtain a separate commercial license from Mistral.mistral
For developers, the practical distinction is straightforward: a company can build a product on top of Le Chonk and run it privately, but cannot simply turn the model into a competing hosted API without negotiating commercial terms.
What developers get
Mistral positions Large 4 as a hybrid instruct-and-reasoning model for coding, agentic workflows, multimodal understanding, and cybersecurity. The company says it was trained from scratch on European infrastructure and supports more than 160 languages.mistral
The model supports three reasoning modes:
| Mode | Recommended use |
|---|---|
| Low | Simple tasks; replaces the previous “None” reasoning mode |
| Medium | Everyday tasks |
| High | Complex or high-precision work requiring deeper reasoning |
High is the default mode.mistral
Mistral also lists support for structured outputs, function calling, document question answering, batching, agents, conversations, and built-in tools through its API.docs.mistral
Coding performance: strong claims, mixed evidence
Mistral says additional reinforcement learning after the October 6 preview substantially improved coding performance. The company reports that Mistral Large 4 now leads its listed open-weight competitors on Cognition’s FrontierCode 2.0 benchmark with a score of 19.7%.mistral
On the Artificial Analysis Coding Agent Index, Mistral says the model scores 48%, effectively tying GLM-5.3 while using 47% fewer tokens per task on average. On Terminal-Bench 4.0, Mistral says the score rose from 28% in preview to 41%, while output tokens fell by 30%.mistral
These are important claims for teams building coding agents, because token consumption directly affects latency and inference cost. But they should be read carefully: several figures are either Mistral-reported or depend on benchmark configurations, and Mistral says a full technical report is still forthcoming.mistral
Independent analysis published before the weights release had also found that Le Chonk trailed some leading Chinese open models on broader capability indexes, even as it offered a distinct advantage in European deployment and self-hosting. That suggests developers should benchmark the model on their own repositories, tools, and workloads rather than relying solely on vendor results.deepinfra
Cybersecurity: capability and risk
Cybersecurity is the most consequential—and sensitive—part of the release. Mistral says Large 4 scores 51% on the Artificial Analysis Cyber Index, which evaluates vulnerability discovery, reproduction, validation, and patching without breaking existing functionality.mistral
The company also says the model refuses 96.6% of malicious cyber requests across selected safety benchmarks, which Mistral describes as the highest refusal rate it measures among open models. It adds that the model resisted 97.1% of attacks in the B3 Agent Security Benchmark, up from 93.3% during preview.mistral
Mistral acknowledges residual risk. In its announcement, the company said live-traffic monitoring found that while the model refused black-box attacks against live targets, it still saw “potential for uplift to conduct real attacks.” Mistral said it is modifying guardrails and will continue updating them.mistral
The company also said enterprise design partners, including Orange Cyberdefense, have preview access to a specialized variant optimized for penetration testing, vulnerability assessment, and red-teaming. Access is restricted to vetted partners and controlled environments involving owned or explicitly authorized systems.mistral
Sovereignty and enterprise deployment
A central part of Mistral’s pitch is sovereign AI: the ability for organizations to control where a model runs, who can access it, and how it is customized. Open weights allow enterprises to deploy the model behind their own firewall, retain data locally, audit the system, and adapt it to proprietary policies or security data.mistral
That positioning is particularly relevant for European organizations. Mistral says Large 4 was developed and trained in Europe, and the release arrives amid growing demand for AI systems that can support data-residency, procurement, and regulatory requirements.mistral
Industry partners cited by Mistral include Booz Allen Hamilton, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, Orange Cyberdefense, Rubrik, Zscaler, ABN AMRO, Belfius, Crédit Agricole, and Microsoft. These statements describe partnerships, evaluations, or support arrangements; they are not independent benchmark validations of every Mistral claim.mistral
Pricing and cost considerations
Mistral’s hosted API pricing for Mistral Large 4 is:
$0.70 per million input tokens
$0.07 per million cached input tokens
$2.10 per million output tokensmistral
For self-hosting, the economics depend on hardware, quantization, serving stack, utilization, and how many tokens an agent consumes. Mistral’s claim of improved token efficiency could reduce costs for coding agents, but teams should measure real-world output volume before assuming savings.mistral
What developers should do next
Developers evaluating Le Chonk should focus on practical testing rather than headline benchmark scores.
Test agentic coding on real repositories. Measure merge-worthy changes, test pass rates, tool-call reliability, and token use.
Benchmark security tasks in isolated environments. Use authorized codebases and controlled sandboxes, especially for vulnerability triage or patch generation.
Compare reasoning modes. Low, Medium, and High may produce materially different cost and quality trade-offs.
Check the license before productizing. Self-hosting and internal use are permitted, but hosted resale requires a separate commercial license.mistral
Validate guardrails independently. Mistral’s safety results are encouraging, but organizations deploying agents should still apply their own policy controls, monitoring, and least-privilege permissions.
What to watch
Mistral says a full technical report is coming, which should clarify architecture, training methodology, benchmarks, and safety evaluation details.mistral
The company also says Mistral Large 4 will serve as the foundation for future specialized and optimized models, extending its broader open-weight strategy.mistral
For now, Le Chonk’s release gives developers a rare combination: a very large multimodal model, downloadable weights, European training lineage, self-hosting options, and a license that permits broad commercial use. Whether it becomes a default choice for coding and security teams will depend less on its trillion-parameter scale than on real-world reliability, cost, safety, and independent evaluation.
