HeyHomees Lab Notes · No. 13 · Mistral · Published October 30, 2026
By HeyHomees Newsroom ·
In 30 seconds
- A 1 trillion-parameter MoE model with 49B active parameters and a 1M token context window.
- Multimodal (text + image) and fluent in 160+ languages, strong on cybersecurity and coding.
- Released under the Mistral Ecosystem Open-Weight License; runs on NVIDIA Grace Blackwell GPUs or via vLLM.
What is it?
Mistral Large 4 (Le Chonk) is Mistral's largest open-weight model, a multimodal Mixture-of-Experts system that accepts text and image inputs and generates text. It has about 1 trillion total parameters and 49B active parameters per token.
How does it work?
The model uses a granular MoE architecture with 128 routed experts, selecting 4 active experts per token, plus one shared expert. It includes a Pixtral-style vision encoder for variable-size images and supports three configurable reasoning effort levels (low, medium, high).
How well does it do?
Mistral reports 82% on the Artificial Analysis Cyber Index, 48% on the Artificial Analysis Coding Agent Index (tied with GLM-5.3), and a score of 26 on the Artificial Analysis Intelligence Index v5, the highest among western open-weight models.
What does it need to run?
Training used 3,800 NVIDIA Grace Blackwell GPUs; inference is recommended with vLLM (or SGLang, transformers, llama.cpp) on similar hardware. A quantized NVFP4 checkpoint is available for Blackwell GPUs with reduced memory.
By the numbers
- Total parameters: 1 trillion
- Active parameters per token: 49 B
- Context window: 1 M tokens
- Supported languages: 160+ languages
- Cybersecurity benchmark: 82%
- Coding agent index: 48%
- Global intelligence score: 26
- Number of routed experts: 128
The honest read
The reported 82% cyber index and 48% coding index are self-reported by Mistral and reflect their internal Artificial Analysis benchmarks; no external third-party evaluations are provided. The 26 intelligence score is also from Mistral's own Artificial Analysis Intelligence Index, which compares only a limited set of open-weight models and does not include closed-source baselines. While the model boasts a 1M token context and multimodal capabilities, performance on real-world tasks beyond the cited benchmarks is not quantified in the sources. Mistral Large 4’s parameter counts (1 trillion total, 49B active) describe the size of the MoE architecture, but the active-parameter figure depends on the token-wise routing of experts, so raw FLOP counts or latency metrics are not given. The model’s strengths in cybersecurity, coding, and visual grounding are highlighted, yet the sources do not provide comparative latency, memory usage, or cost figures for inference.
What we don't know yet
- Exact inference latency and cost per token on typical hardware
- Performance on non-cybersecurity multilingual tasks beyond the cited benchmarks
- Availability of a non-GPU (CPU-only) deployment option
Sources
- Model card, Hugging Face (primary)
- Hugging Face Blog
Researched and drafted with AI from the sources above; every figure is checked against them by code before it is published, and a line that fails is dropped.