Mistral Small 4
2026-03-16 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI launches Mistral Small 4, an open source Apache 2.0 model with 119 billion parameters (6 billion active) that combines instruction following, reasoning, vision and coding agents in a single model.
What changes
- First Mistral model to unify Magistral (reasoning), Pixtral (multimodal) and Devstral (agentic coding).
- reasoning_effort parameter (none or high) to choose between fast answers and deep reasoning.
- Mixture of experts (MoE) architecture: 128 experts, 4 active per token.
- End-to-end latency reduced by 40% and throughput multiplied by 3 compared with Mistral Small 3 (optimized configurations).
### Specifications - Parameters: 119B total, 6B active per token (8B with embeddings and output layer). - Context: 256k tokens. Text and image inputs. Maximum output and knowledge cutoff: not disclosed. - License: Apache 2.0. - Availability: Mistral API, AI Studio, Hugging Face, NVIDIA NIM, vLLM, llama.cpp, SGLang, Transformers. - Minimum infrastructure: 4 HGX H100 GPUs, 2 HGX H200 or 1 DGX B200. ### Pricing | Item | USD per million tokens | |---|---| | Input | 0.15 | | Output | 0.60 | These prices come from a third-party guide (Serenities AI), not from the official post, which does not mention them. ### Benchmarks | Measure | Result | |---|---| | AA LCR (with reasoning) | 0.72 for 1.6K characters of output | | LiveCodeBench | higher than GPT-OSS 120B with 20% less output | Scores published by the vendor (shown in charts, partial values). ### Key takeaways A single open model covers chat, reasoning, vision and code, which simplifies self-hosting for companies that require control over their data.
Links
Tags : LLM, Mistral, Small, Open Source, Multimodal, Raisonnement