About This Video
A 21-minute review from Matthew Berman of Mistral Large 4, nicknamed "Le Chonk" after a long-running community meme. He covers the specs, pricing, benchmarks, and the European sovereignty angle, then spends the last stretch on his complaints about how hard open-source models are to use in agent harnesses. Timestamps below are approximate (placed proportionally from the auto-captions).
Summary
Mistral released a public preview of a frontier model built completely in Europe, trained from scratch rather than on top of a Chinese open-source base. It is a one trillion parameter mixture-of-experts model with 49 billion active parameters, priced at $1.36 per million input tokens and $4.18 per million output tokens. Berman finds it very competitive among open-weight models, especially on cybersecurity, but notes that the US closed-source frontier is still at least six months ahead and that a Chinese open-source model still beats it almost every time.
Key Facts
- Size: one trillion total parameters, 49 billion active; hybrid instruct and reasoning, natively multimodal input.
- Weights: not released yet, but promised by the end of the month. Meanwhile it is being red teamed with cybersecurity leaders, vetted partners, and state authorities who get a reduced-moderation version.
- Training: from scratch on 3,800 Nvidia Grace Blackwell GPUs in Mistral's own European data center, which also serves the public preview.
- Context window: 500K tokens, versus the roughly 1M standard on most competitors.
- Speed: 116 tokens per second, versus 227 for DeepSeek V4.1 Flash.
- Availability: API key from Mistral; not yet in the Le Chat app.
Benchmarks Mentioned
- Cyber Gym: 82, the top score among open-source models, including the Chinese ones.
- AA Cyber Index: 50, tied for first with GLM 5.3 Flash. Mistral says it ranks among the top five models globally on cybersecurity.
- Automation bench: 59.9, versus 62.2 for GLM 5.3 and 58.3 for Kimi K3, a big jump from Mistral Medium 3.5.
- Terminal Bench: second place among open models, with GLM 5.3 at 40 and Kimi K3 at 22.
- Artificial Analysis Intelligence Index: 38, which puts it 25th of 25 models, with Anthropic models holding the top six spots and GPT-6 Astra at 53.
Berman's Main Points
- Sovereignty matters. The model was conceived, trained, and is served from Europe. Until now it has been the US with the top closed models and China with the top open ones, and he is glad to see a third competitor.
- Closed labs are pacing the frontier. GPT 6.1 Soul and Opus 5.5 are fast, cheaper, and top the charts, so much of what drew people to open source (price, speed) is now available from frontier labs. Open source still offers ownership, privacy, and zero data retention.
- Open weights compound. Once many people have the model, they will squeeze out more performance and tune it for narrow use cases where it could beat the closed models.
- Usability is the barrier. Plugging an open model into an agent harness is not plug and play. Open source works best today for teams with expertise operating at scale.
The Rubik's Cube Test
Berman used the standard "build a Rubik's cube simulation" prompt in OpenCode. At high thinking effort the model burned through tokens until it hit about 32,000 and stopped. After he raised the context limit and turned off thinking, the cube rendered and spun, but the scramble animation first did nothing, and a later version lost the cube's colors. He blames the model-and-harness pairing rather than the model, and says Cursor is the exception that makes other companies' models work well.
Reflect & Discuss
- When does full control (hosting, guardrails, privacy) justify a weaker and more expensive model than the closed frontier?
- Berman argues the harness matters as much as the model. What would make an open model plug and play for non-experts?
- Does it matter where a foundation model is built and served, and for which kinds of organizations?