Mistral Large 4 Drops: A 1T Parameter Open-Weight Beast for Enterprise
Mistral just previewed a 1-trillion parameter model that beats US open-weight rivals and refuses to censor legitimate cybersecurity work.
Models · Source: Hacker News
What happened
Mistral just launched the public preview of Mistral Large 4. They call it le Chonk internally. It is a massive 1-trillion parameter natively multimodal model. It operates with 49 billion active parameters. The model demonstrates exceptional performance across coding, agentic workflows, and multimodal understanding. Mistral generates roughly 33 billion tokens per day during its reinforcement learning training. About 16 billion of those are trainable completion tokens. This massive scale allows the model to adapt rapidly. The weights drop at the end of this month.
The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs. Mistral used its own datacenters in Europe for this run. It is built specifically for AI sovereignty. Mistral claims it significantly outperforms any open-weight model developed in the US or Europe. A significant share of the training data was multilingual. It spans more than 160 languages. This includes every official language of the European Union. Mistral worked closely with leading enterprises in finance, engineering, and manufacturing to train the model.
ML4 is currently available via the Mistral Studio API. Early red-teaming is happening right now. Mistral is working with cybersecurity leaders, vetted partners, and state authorities. These groups get a version with reduced moderation to test expanded cyber capabilities. The final release will be available across multiple regions worldwide.
Key facts
- 1 trillion — Total parameters in Mistral Large 4
- 49 billion — Active parameters during inference
- 3,800 — NVIDIA Grace Blackwell GPUs used for training
- 82% — Score on AA Cyber Index vulnerability reproduction test
- 33 billion — Tokens generated per day during RL training
Why it matters
This changes the game for enterprise AI deployment. Closed models from competitors like Anthropic and OpenAI often block legitimate vulnerability research. Their strict safety filters get in the way of actual work. Mistral Large 4 offers open weights and self-deployment. You get top-tier capabilities without a provider shutting down your incident response mid-investigation. ML4 ranks among the top five models globally on the Artificial Analysis Cyber Index. It scores 82 percent on a test that asks a model to reproduce a real vulnerability and patch it. It also solves 93 percent of challenges in Cybench. These numbers prove it can handle real-world security tasks.
The second-order effect is a massive shift in sovereign AI capabilities. European and global enterprises can now run a frontier model entirely on-premise or in private clouds. This removes reliance on US-based digital service providers. It ensures compliance with local data laws. The model also brings strong scientific capabilities. It can generate a full Hartree-Fock simulation in one shot. Researchers can focus on hard questions rather than plumbing. This forces closed-model providers to rethink their heavy-handed refusal policies. If they do not adapt, they will lose lucrative enterprise contracts to open-weight alternatives.
For builders
Self hosted cybersecurity workflows
Security teams can deploy ML4 on-premise to analyze malware and write detection rules. You avoid the massive risk of closed models refusing your prompts during an active incident. Enterprise security vendors win big here, while closed API providers lose ground.
Agentic coding and automation
ML4 scores high on DeepSWE and AutomationBench for business workflows. Builders can integrate it into complex terminal operations and apps like Salesforce or Sheets. Companies paying for expensive closed-model API calls can cut costs by hosting this themselves.
Visual grounding for complex documents
The model surpasses GPT-6 Astra in visual grounding tasks like Dense 200. You can build apps that inspect gigapixel satellite imagery or verify mechanical parts in technical drawings. Engineering and manufacturing firms will pay a premium for this level of accuracy.
My take
Mistral is proving that open weights are the only viable path for serious enterprise security. Closed models like Claude Opus 5.5 and GPT-6 Astra are useless to me if they refuse to do the actual work. I build AI products to solve hard problems, not to get lectured by an overly sensitive safety filter.
Original reporting: Hacker News. This is my rewrite and opinion.