Introducing Mistral Small 4
Mistral AI
Mistral just dropped Small 4, one open model that reasons, codes, and reads images instead of needing three separate ones. It matches GPT-OSS 120B on benchmarks while writing way shorter answers, which means cheaper and faster.
Based on reporting by Mistral AI — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Mistral has a habit of shipping models faster than most people can keep their release names straight, and today's Small 4 is the company folding three of those names into one. Instead of picking between Magistral for reasoning, Pixtral for images, or Devstral for coding agents, you now get all three in a single Apache 2.0 checkpoint. That's the pitch, anyway: one model, configurable behavior, no more juggling.
Under the hood it's a mixture-of-experts setup with 128 experts and only 4 active per token, landing at 119 billion total parameters but just 6 billion active (8 billion if you count embeddings). That sparsity is the whole trick — you get a model that behaves like something much smaller at inference time while still having the capacity of a much bigger one. Add a 256k context window and native image input, and Small 4 is clearly built to be the default choice rather than a niche specialist.
The headline number, though, is efficiency. Mistral says Small 4 cuts end-to-end completion time by 40% in a latency-tuned setup and handles 3x the requests per second compared to Small 3 when tuned for throughput. On benchmarks like AA LCR, it scores 0.72 using around 1.6K characters of output, while comparable Qwen models need 3.5 to 4 times more text to hit similar accuracy. On LiveCodeBench it beats GPT-OSS 120B while writing 20% less. Shorter answers at equal or better accuracy is the kind of unglamorous win that actually shows up on your cloud bill.
There's a new reasoning_effort dial too, letting you flip between fast, low-latency chat responses and Magistral-style deep reasoning without swapping models. Mistral is also leaning hard into partnerships here, joining the NVIDIA Nemotron Coalition, shipping day-zero as an NVIDIA NIM, and getting inference optimized for vLLM and SGLang. The weights are on Hugging Face, it runs on llama.cpp and Transformers already, and Mistral is positioning this as infrastructure you can actually self-host, not just call through an API.
My take — AI-written commentary, not fact-checked reporting
This is the sensible move — collapsing three specialist models into one general-purpose model is exactly what enterprises actually want, because nobody wants to maintain three inference pipelines when one will do. The output-length efficiency numbers matter more than the accuracy scores here, since token count is the real cost driver at scale, and Mistral clearly understands that better than most labs still chasing benchmark leaderboards. Apache 2.0 on something this capable is also a genuine shot at the closed-model pricing model, and I'd bet this becomes the default self-hosted pick for a lot of teams within a quarter.
Read more about this at: Mistral AI