Mistral releases Le Chonk, a 1.05 trillion parameter open-weight AI model for coding
The divide between closed-source AI giants and the open-weight community has narrowed significantly with the release of Mistral Large 4. Nicknamed "Le Chonk" due to its massive scale, the model represents a strategic push by the French AI lab to provide a European alternative to the proprietary systems dominated by American and Chinese firms.
Following a public preview on October 6, 2026, Mistral officially released the weights for Mistral Large 4 on October 29. The model is a natively multimodal mixture-of-experts (MoE) system that allows users to input both text and images. To achieve this, Mistral utilized a 1.6-billion-parameter Pixtral-style vision encoder designed to handle various aspect ratios and image sizes. The sheer scale of the model is reflected in its architecture:
- Total parameters: Approximately 1.05 trillion
- Active parameters per token: Roughly 52 billion
- Expert configuration: 128 routed experts (four active per token, plus one shared expert)
- Context window: 1 million tokens
This massive footprint was developed using a relatively lean compute budget compared to its US rivals. Mistral trained the model from scratch over two months using between 3,800 and 4,000 Nvidia Grace Blackwell GPUs within its own European data centers. This stands in stark contrast to the resources used by proprietary labs; for instance, Nvidia CEO Jensen Huang noted that OpenAI utilized roughly 100,000 GPUs for GPT-6 Astra.
Coding and Reasoning Performance
The release of the weights comes alongside updated benchmark scores that position Le Chonk as a premier open-weight model for coding. On the Terminal-Bench 4.0 agentic coding benchmark, the model's score rose to 41%, up from 28% in the preview version, allowing it to edge out Opencode’s GLM-5.3. While it remains effectively tied with Codex GPT-6.1 Terra and slightly behind GLM-5.3 on the overall Artificial Analysis Coding Agent Index with a score of 48%, Mistral claims the model is significantly more efficient, using an average of 47% fewer tokens per task than GLM-5.3.
To manage the costs and performance associated with such a large model, Mistral introduced three distinct reasoning modes: low, medium, and high. The "high" mode serves as the default for complex tasks requiring deep reasoning, while "low" replaces the previous "none" mode for simpler operations. This flexibility is paired with a DSpark speculative-decoding draft head, which Mistral reports can increase throughput by up to 1.9x compared to serving the target model alone.
Cybersecurity and Sovereign AI
Beyond general coding, Mistral has positioned Mistral Large 4 as a critical tool for defensive cybersecurity. The model scored 51% on the Artificial Analysis Cyber Index, demonstrating a high proficiency in discovering, reproducing, and patching memory-safety vulnerabilities in C/C++ open-source projects. Mistral emphasizes that this focus on "sovereign AI" allows governments and enterprises to maintain control over their data and security tools without relying on proprietary labs that could potentially rescind access.
This defensive capability is balanced by strict guardrails. Mistral reported a 96.6% refusal rate for malicious cyber requests across benchmarks like JailbreakBench and AgentHarm. However, the company acknowledged that it observed potential for the model to be used in real-world attacks during live traffic monitoring following the preview. In response, Mistral is modifying its guardrails and providing specialized variants for penetration testing to vetted partners, such as Orange Cyberdefense, under strict contractual safeguards.
The deployment of Le Chonk is supported by a modified MIT open-weight license via Hugging Face. Under the Mistral Ecosystem Open-Weight License, the model is free to modify and commercialize, though businesses providing Model-as-a-Service offerings must obtain a separate commercial license. For those preferring API access, Mistral has priced the model at $0.70 per million input tokens and $2.10 per million output tokens.
This aggressive release is backed by a massive influx of capital. Mistral recently raised 3 billion euros in funding led by Samsung Electronics and the EU-backed Scaleup Europe Fund, bringing the company's post-money valuation to over 21 billion euros. CEO Arthur Mensch has indicated that this funding will be used to expand Mistral's own data center footprint, reducing reliance on external compute and ensuring the lab can continue to train larger, faster models independently.
The arrival of Mistral Large 4 signals a shift in the AI landscape where "open" no longer means "underpowered." By matching the coding and security capabilities of closed-source frontier models while maintaining a commitment to open weights, Mistral is challenging the prevailing industry assumption that state-of-the-art performance requires a closed ecosystem. The success of Le Chonk will likely determine if European "sovereign AI" can move from a political ambition to a technical reality.
