Report · mistral.ai
We're releasing the weights of Mistral Large 4
mistral.ai
Source published October 29, 2026
Summary: 2 min read
81/100
Summary
Mistral has released the weights for Mistral Large 4, also called Le Chonk, a 1-trillion-parameter natively multimodal mixture-of-experts model that had entered public preview earlier in the month. The open-weight model is designed for coding, agentic workflows, defensive cybersecurity and enterprise tasks, and offers low, medium and high reasoning modes to balance cost and performance. Mistral says the model was trained from scratch on European infrastructure and can be inspected, customized and self-hosted. Mistral reports that three additional weeks of reinforcement learning raised its FrontierCode 2.0 score to 19.7% and its Terminal-Bench 4.0 result by 13 percentage points to 41%. On the Artificial Analysis Coding Agent Index, it scored 48% while using 47% fewer tokens per task on average than GLM-5.3, according to Mistral. The company also reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% average refusal rate for malicious cybersecurity requests. A specialized cybersecurity version is being previewed with vetted enterprise partners for authorized testing. Mistral Large 4 is available through Mistral Studio, the Mistral API, OpenRouter, Hugging Face, Microsoft Foundry and NVIDIA-hosted endpoints. API pricing starts at $0.70 per million input tokens and $2.10 per million output tokens. Its modified MIT-style license permits use, modification, fine-tuning and commercialization, except Model-as-a-Service providers require a separate commercial license.
Key Takeaways
- Mistral released the open weights of its 1-trillion-parameter, natively multimodal Mistral Large 4 model, enabling organizations to self-host and customize it.
- Mistral says Mistral Large 4 scored 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task on average than GLM-5.3.
- Mistral reports that the model scored 51% on the Artificial Analysis Cyber Index and refused an average of 96.6% of malicious cybersecurity requests.
- Mistral Large 4 supports low, medium and high reasoning modes, with high set as the default for tasks requiring greater precision and more thorough reasoning.
- The model costs $0.70 per million input tokens and $2.10 per million output tokens through Mistral's API, while Model-as-a-Service businesses need a separate commercial license.
What the discussion said
The thread spent less time celebrating a new model than arguing over what its release actually permits. The weights are downloadable, but the serving restriction means commercial model-as-a-service use needs a separate deal. Many readers see that as source-available rather than genuinely open, and worry that frontier AI is becoming dependent on lab-controlled licenses because training costs are too high for a durable community alternative. Others say open-weight has always meant public access to weights, not unrestricted commercial rights, and regard the practical freedom to run internally or self-host as the important part. On the product itself, sentiment was guarded. The lower API price made several readers reassess it, especially for European companies seeking regional data handling, a broader API platform, and access to third-party models. Yet hands-on reports describe it as merely competent: useful for structured coding work, but prone to hallucinations or ignored instructions, slow under reasoning, and behind leading Chinese and US open-weight competitors. Self-hosting also remains a serious hardware commitment; even a four-H200 machine reportedly requires aggressive quantization or offloading. Readers were particularly unconvinced by the cyber-safety refusal statistic, arguing that refusal rates without benign-request utility are uninformative and that a self-hosted model can be stripped of safeguards readily.
Where opinion split
The sharpest fight is whether Mistral can call this an open-weight release. Critics argue that a commercial serving restriction defeats the point of openness and leaves users exposed to vendor permission. Defenders argue that published weights still deliver the core benefit—local and internal use—and that most users already expect conditions on commercial redistribution or hosted access.
Coverage & discussions
Hacker News · mistral.ai
Community discussions
Community Sentiment
Positives
- The reduced API pricing makes a large frontier-adjacent model materially more plausible for production, particularly for teams that value European hosting and data residency.
- Mistral’s expanded platform—third-party models, newer API features, and regional service—gives EU businesses a practical alternative to routing sensitive workloads through US providers.
- Public weights create an important self-hosted option for regulated organizations that cannot deploy Chinese models, even if the model demands heavy quantization and substantial GPU memory.
- Several practical users find it capable for disciplined coding and document-tool workflows, where well-specified tickets matter more than one-shot autonomous generation.
Concerns
- The commercial serving clause turns a supposedly open-weight release into vendor-controlled weight availability, undermining unrestricted experimentation and downstream AI businesses.
- Commenters doubt the model is competitive on capability or cost, citing disappointing benchmarks and real-world results behind stronger Chinese and US open-weight models.
- Reports of hallucinations, ignored project instructions, and slow reasoning make the model hard to justify when higher-performing alternatives are readily available.
- The touted cyber refusal rate is dismissed as a safety headline without evidence that legitimate security requests still work, especially since self-hosters can readily remove guardrails.
