Today for AI
HOT RADAR
llmHEAT 45.9°

AI CURATED HOT STORY · 10/29/2026

Mistral Releases Open-Source Trillion-Parameter Model Le Chonk

3 reports archived3 independent sourcesupdated 10/29/2026, 23:25:27
Synthesis & Latest Updates
3 Sources Cross-Validated

Mistral AI has officially open-sourced the weights for its flagship model, Mistral Large 4 (codenamed Le Chonk), marking a significant milestone in European sovereign AI. The model features approximately 1 trillion total parameters and 52 billion active parameters within a Mixture-of-Experts (MoE) architecture, supporting multimodal inputs. Weights, GGUF quantizations, and inference endpoints are now available on Hugging Face Hub for immediate local deployment or fine-tuning.

Cross-referenced reports highlight Le Chonk's leading performance in agentic coding and cybersecurity defense, boasting a 96.6% refusal rate for harmful requests and efficient inference speeds of 161 tokens/s. While its general intelligence index trails some Chinese open-source models and top closed systems, its three inference modes allow developers to balance cost and performance effectively under a modified MIT license.

This release strengthens Mistral's position in the high-value open-source ecosystem, offering enterprises a secure, low-latency alternative with data sovereignty advantages. By prioritizing safety and coding capabilities, Le Chonk aims to fill gaps in the market for specialized, high-performance open models.

LATEST/Hugging Face fully supports Le Chonk downloads and deployment, with the community rapidly generating quantized versions. Developers are leveraging its three inference modes to optimize cost-efficiency, focusing on real-world performance in code generation and security filtering.

Hourly heat curve

2 fully observed hours
Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 3 reports · Latest first
  1. The DecoderT2·78 pts

    Mistral Releases 'Le Chonk', a Trillion-Parameter Open Coding Model

    • Mistral Large 4 ('Le Chonk') releases open weights with ~1T total and 52B active parameters under a modified MIT license.
    • Strong coding performance, but scores 26 on the Intelligence Index, trailing Chinese open models like DeepSeek V4.1 Flash.
    • Focuses on efficiency and safety: 161 tokens/sec speed, low API costs, and a 96.6% refusal rate for harmful cyber requests.
  3. Hugging Face BlogT1·85 pts

    Mistral Large 4 Launches on Hugging Face with Multimodal MoE Architecture

    • Mistral Large 4 utilizes a Mixture-of-Experts (MoE) architecture supporting multimodal inputs and multilingual processing.
    • Hugging Face has simultaneously released official weights, GGUF quantized formats, and inference endpoint services.
    • Developers can directly access models on the Hub for local deployment, fine-tuning, or community sharing.