Mistral AI Unveils 1T Parameter 'Le Chonk' Model
Mistral AI has introduced Mistral Large 4 (ML4), a 1.05 trillion-parameter multimodal model that the French lab is positioning as a direct competitor to both American and Chinese frontier models. Nicknamed Le Chonk, the model utilizes a Mixture of Experts architecture and is currently available via a public guardrail endpoint, with full open-weight release scheduled for October 27, 2026. The launch follows a €3 billion Series D funding round led by Samsung, which valued the company at approximately €21 billion.
Competitive Positioning and Pricing
Mistral is aggressively targeting the enterprise market by undercutting the pricing floor established by proprietary incumbents. ML4 is priced at $0.68 per million tokens for input and $2.09 for output, a significant reduction compared to the $2/$10 baseline often associated with models like OpenAI's GPT-6.1 Sol and Google's Gemini 4 Argon. By offering an open-weight alternative, Mistral aims to capture market share from both closed-source systems and smaller open-source players, while maintaining its status as a frontier lab.
Infrastructure and Sovereignty Paradox
The development of ML4 highlights a complex reality for European AI sovereignty. While the model was trained entirely on Mistral's own infrastructure, utilizing 4,000 Nvidia Grace Blackwell GPUs hosted in European data centers, the company remains dependent on American hardware supply chains. Mistral VP of Science Pierre Stock noted that the model was trained using two to three times less compute than its Chinese competitors, emphasizing an efficiency-focused training strategy.
Mistral co-founder Guillaume Lample disclosed that the model's post-training reinforcement learning run remains in active execution with substantial headroom remaining ahead of the late October weights release.
Benchmarks and Verification
Mistral has reported strong performance metrics, including 62% on DeepSWE and 49.8% on the Coding Agent Index. However, these figures remain vendor-reported and lack third-party verification, a point of friction for enterprise decision-makers. The company plans to work with cybersecurity specialists and public authorities to assess the model's behavior under real-world conditions before the weights are published. Mistral intends for ML4 to be used in high-stakes applications such as cybersecurity, finance, and industrial chip design, positioning the model as a tool for defense against malicious actors.
How to compare a trillion-parameter model
The total parameter count in a mixture-of-experts system does not describe how much computation every request uses. Only a portion of the network may be activated for each token, so latency, memory requirements and real-world price depend on deployment choices. Mistral's reported benchmark scores are useful starting points, not independent proof that one model is better for every task. Enterprises considering the promised open-weight release should also assess the actual licence, availability of weights, safety documentation and reproducible evaluations. A European training location may support data-sovereignty goals, but it does not eliminate reliance on globally sourced accelerator hardware.