The Pulse
Mistral Large 4: Model Details, Variants and Deployment Options
Hugging Face’s guide and Mistral’s model card describe Large 4’s architecture, capabilities, variants and ways to run it. The documents also set out the model’s license and report results on coding and cybersecurity evaluations.
AI.info Team · Written with AI · fact-checked ·
“Mistral Large 4 (aka Le Chonk) is a multimodal, multilingual, Mixture-of-Experts (MoE) model from Mistral.”
Merve Noyan and the other Hugging Face authors write
Hugging Face’s guide, published October 29, 2026, describes Mistral Large 4 as a model that takes text and images as input and generates text. Mistral’s model card lists about 1 trillion total parameters, 49 billion active per token and 52 billion including embeddings and output layers. Both documents describe a context window of up to 1 million tokens.
Architecture and capabilities
Mistral’s model card describes a mixture-of-experts architecture with 128 routed experts, four active for each token, and one shared expert. It says the model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s European data centers. The card characterizes Large 4 as a single set of weights for instruction following, reasoning, coding and agentic workflows, and says it is natively fluent in more than 160 languages.
Hugging Face’s guide provides further architectural details: the text decoder has 72 layers, 64 attention heads and a hidden dimension of 6,144. Its attention combines Multi-head Latent Attention with a DeepSeek Sparse Attention indexer, which selects up to 2,048 key positions for each query. The guide also describes a 1.6-billion-parameter Pixtral vision encoder. Mistral’s model card says the Pixtral-style encoder is trained to handle variable image sizes and aspect ratios.
Variants and inference
The guide lists two official variants, FP8 and NVFP4, and notes that GGUF versions are also available from Unsloth and the llama.cpp team. It describes the NVFP4 checkpoint as having a smaller memory footprint and targeting NVIDIA Blackwell GPUs. Mistral’s model card says a quantized NVFP4 checkpoint supports vLLM and SGLang.
The Hugging Face guide documents support for Transformers, vLLM and llama.cpp, along with remote deployment through Hugging Face Inference Endpoints and serverless inference through Inference Providers. Its Transformers example uses eight processes. The vLLM instructions serve the NVFP4 checkpoint and set tensor and decode-context parallelism to eight. The model card separately includes examples for serving the checkpoint with vLLM and SGLang, as well as using Transformers.
The guide says reasoning effort can be set to low, medium or high, with high as the default. It also says reasoning cannot be disabled through the documented setting: setting the effort to none still uses high. The model card likewise describes reasoning effort as configurable per request and lists the same three settings.
License and published evaluations
Mistral’s model card says the weights are released under the Mistral Ecosystem Open-Weight License, a modified version of the MIT license. Commercial and non-commercial use are allowed, but operating a Model-as-a-Service business requires a separate commercial license from Mistral.
Hugging Face’s guide reports results across coding and cybersecurity evaluations. It gives Large 4 a score of 82% on the Artificial Analysis Cyber Index vulnerability-reproduction and patching test, and reports 93% on Cybench, a suite of 40 security challenges. For coding, the guide lists scores of 61.7% on DeepSWE v1.1 and 28.3% on Terminal-Bench 4. These are results on specific evaluations, not a general guarantee of performance.
Together, the guide and model card describe a model that accepts text and image inputs, with official FP8 and NVFP4 variants and several documented inference options. The NVFP4 checkpoint is specifically described as targeting Blackwell GPUs; the license also sets a separate requirement for businesses operating a hosted Model-as-a-Service offering.