Mistral Releases Le Chonk Open-Weight AI Model for European Sovereign Control
French AI firm Mistral has opened the weights of its massive Mistral Large 4 model, known as Le Chonk, giving organizations a new self-hosted option for coding and security tasks.
European artificial intelligence development has reached a new milestone as French firm Mistral has made the weights of its largest AI model publicly available. Frequently referred to by the nickname Le Chonk, the large-scale model can now be downloaded, operated on local infrastructure, and customized by organizations. Initially previewed via an API in October, the model underwent further training to boost its reasoning capabilities, cybersecurity functions, and software engineering performance.
This open-weight release provides teams with multiple ways to interact with the technology. Organizations can utilize it via managed API setups, choose hosted deployments, or run the system internally under Mistral's specific licensing terms. For instance, a European software development team building a coding assistant can carefully evaluate where their code is processed and who operates the underlying infrastructure before finalizing their technical architecture.
Performance and Competitive Landscape
Mistral's decision places its model directly into a competitive arena alongside prominent open-weight alternatives from Asia, including Z.ai's GLM-5.3, Moonshot AI's Kimi K3, and DeepSeek V4.1 Flash. Company-reported metrics show Le Chonk scoring 48% on the Artificial Analysis Coding Agent Index, matching closely with GLM-5.3 while reportedly consuming 47% fewer tokens per task during those evaluations. Additionally, the system reached 41% on Terminal-Bench 4.0, marking a notable improvement over its initial preview phase.
Independent human evaluations, such as Surge Code blind tests, reveal a nuanced picture. While evaluators often preferred Le Chonk's outputs over competing models from Chinese labs, American models like Claude Opus 5.5 still ranked ahead in overall preference. Real-world speeds also vary significantly based on infrastructure. Hugging Face deployment tests using eight NVIDIA H200 GPUs recorded roughly 65 generated tokens per second with FP8 weights, while a quantized version running on an Apple Silicon Ultra with 512 GB of unified memory achieved about 15 tokens per second.
What it means for developers
For engineering teams choosing a coding assistant, self-hosting Le Chonk offers significantly more control over access controls, security policies, and data processing environments compared to standard third-party APIs. However, this autonomy introduces hardware, operational, and licensing responsibilities, as businesses offering the model as a service require separate commercial agreements.
When exploring different artificial intelligence architectures, developers can try top AI models cheaply through one API at https://apixoai.online. Evaluating any model requires testing it against representative internal tasks—such as bug fixes or test generation—while carefully measuring end-to-end execution times, infrastructure costs, and token efficiency.
Infrastructure and Practical Control
Le Chonk utilizes a mixture-of-experts architecture, totaling roughly a trillion parameters with approximately 52 billion active for any given token. This design reduces processing overhead per step, but the massive scale still demands substantial hardware resources for local execution. While built-in guardrails and defensive security features assist with compliance, modifying deeply ingrained model behaviors requires extra effort and technical capability. Ultimately, the release gives organizations a powerful European-developed alternative, leaving teams to decide whether local infrastructure control outweighs the operational costs.
Source: Mistral Large 4 Gives Europe a New Choice in Open-Weight AI — Towards AI. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
