Mistral’s giant AI model is downloadable now—but hosted services face a license line
Mistral released downloadable weights for its trillion-parameter Large 4 model, giving organizations more control over deployment while requiring a separate commercial license for certain hosted model services.
Mistral Large 4 (Le Chonk) is now open weight
Mistral releases weights for its trillion-parameter AI model, targeting cybersecurity
Welcome Le Chonk: Mistral Large 4 on Hugging Face
Mistral Says Its AI Model ‘Le Chonk’ Has Leapfrogged China in the Open-Weight Race
Mistral Large 4 'Le Chonk' arrives as the top open model outside China with improved coding performance and 63% lower cost-per-task
Downloadable weights change who runs the model
On Oct. 29, Mistral AI released downloadable weights for Large 4, shifting its trillion-parameter system from API preview to a checkpoint companies can run and customize themselves. The company calls it “Le Chonk”; the model takes text and images as input. The weights are available on Hugging Face mistral +1. Hugging Face’s model card documents local deployment, including serving frameworks huggingface.
Self-hosting matters to organizations seeking to keep data and model operations on their own infrastructure. Mistral says three more weeks of training after the preview improved coding and defensive cybersecurity performance; those are company claims, not a guarantee of results in every deployment mistral +1.
“Open” has a service boundary
Large 4 uses Mistral’s “Ecosystem Open-Weight License,” a modified MIT license mistral +1. It allows use, modification, fine-tuning and commercial deployment, but a company or affiliate that operates a model-as-a-service business must secure a separate commercial license before serving the model to third parties. The text has exceptions for certain embedded end-user products and internal use; it is not a blanket ban on commercial use mistral +1.
Self-hosting remains open, while some hosted services require a separate agreement techcrunch +1.
The security pitch cuts both ways
Mistral pitches the model for defensive work such as finding and patching vulnerabilities. It reports refusing 96.6% of malicious cyber requests in its tests, but says monitoring of live traffic showed potential to help conduct real attacks, prompting changes to its guardrails. Selected security partners can test a specialized version under vetting and contractual safeguards mistral.
Those safeguards do not travel as an enforceable promise with weights others can download and modify. WIRED describes a dual-use trade-off: defenders gain a model they can operate independently, but the same availability can help less benign users wired. Local control lets organizations set their own defenses; it does not establish that downstream copies will keep Mistral’s protections techcrunch +1.
Strong coding numbers, not an overall crown
Artificial Analysis says its Intelligence Index combines nine evaluations, making it a defined benchmark rather than a verdict on every workload artificialanalysis. Independent coverage put Large 4 at 26, one point above its preview but below DeepSeek V4.1 Flash and GLM-5.3 at 32 venturebeat +1.
Mistral’s sharper case is coding efficiency: VentureBeat reported a 48 score on the Coding Agent Index, just behind GLM-5.3’s 49, with about 47% fewer tokens per task venturebeat. That is a benchmark comparison, not measured savings for a customer’s own systems; Mistral describes its coding results as leading among open models mistral.