Mistral Drops 1T Parameter Le Chonk Model Targeting Cyber and Agents
Mistral Large 4 brings frontier performance to open weights, letting builders run advanced cyber workflows without API refusals.
Models · Source: Hacker News
What happened
Mistral just launched a public preview of Mistral Large 4, officially codenamed Le Chonk. The natively multimodal model packs one trillion total parameters with 49 billion active parameters. It was trained from scratch on 3800 NVIDIA Grace Blackwell GPUs in European datacenters. You can test the API today on Mistral Studio, and the open weights will drop at the end of October.
The model targets critical enterprise workloads like cybersecurity, finance, and law. Mistral claims it outperforms any open-weight model developed in the US or Europe by a wide margin. It even beats closed models like GPT-6 Astra in specific visual grounding tasks, scoring 42 percent to their 41 percent. It also scores highly on agentic coding benchmarks, beating competitors like DeepSeek V4 Pro and Qwen3.8 Max on the Coding Agent Index.
Mistral is leaning heavily into reinforcement learning at scale to train this model. Their pipeline generates 33 billion tokens per day across 3000 GPUs. This setup allows them to continuously adapt the policy to increasingly complex tasks. The training data was highly multilingual, covering over 160 languages including every official language of the European Union.
Key facts
- 1 trillion — Total parameters in Mistral Large 4
- 49 billion — Active parameters during inference
- 3800 — NVIDIA Grace Blackwell GPUs used for training
- 82% — Score on reproducing and patching real vulnerabilities
- 33 billion — Tokens generated per day during RL training
Why it matters
Builders in highly regulated or sensitive industries finally have a frontier-level model they can run on private infrastructure. Closed APIs from OpenAI and Anthropic often block legitimate cybersecurity workflows due to aggressive safety filters. Claude Opus 5.5 and GPT-6 Astra score near zero on certain cyber tests simply because they refuse the prompt. Mistral Large 4 removes this friction entirely. It gives security teams the autonomy to analyze malware and patch vulnerabilities without losing access mid-incident.
This release shifts the balance of power in enterprise AI sovereignty. Mistral is proving that European infrastructure can train a trillion-parameter model that rivals US tech giants. This makes self-hosted AI a viable default for governments and massive corporations. If organizations can get top-tier performance without sending data to a third party, closed API providers will lose their grip on the most lucrative enterprise contracts.
For builders
Unrestricted cybersecurity workflows
Security engineers can deploy this model locally to reverse-engineer malware and write detection rules. You no longer have to pay closed API providers who might flag and block your legitimate vulnerability research.
Complex agentic automation
Founders building AI agents get a massive upgrade for multi-step tasks. The model excels at navigating spreadsheets, PDFs, and terminal workflows. This makes it cheaper and more reliable to build autonomous knowledge workers for finance and law.
Sovereign deployments for enterprise
Startups selling to the public sector or defense can now offer fully air-gapped AI solutions. You win contracts by guaranteeing that sensitive client data never leaves their servers.
My take
Mistral is doing exactly what open-weight AI needs right now. They are attacking the glaring weakness of closed models by removing overzealous safety filters that break enterprise workflows. If Claude and GPT refuse to do the actual work, builders will flock to the model that just gets it done. I love seeing a European company push the frontier on their own terms.
Original reporting: Hacker News. This is my rewrite and opinion.