Mistral Small 3
A 24B open-weight model tuned for low latency and local deployment
Mistral Small 3 is a 24-billion-parameter language model from Mistral AI, released in January 2025 in both pretrained and instruction-tuned forms under the Apache 2.0 licence. The company positions it as a fast, efficient option for everyday generative AI tasks, competitive with models several times its size and suitable for running privately on a single consumer GPU or laptop.
- 24B parameters, released under the Apache 2.0 licence
- Over 81% on MMLU at around 150 tokens per second, per Mistral AI
- Quantised, it runs on one RTX 4090 or a 32GB MacBook
- Trained without reinforcement learning or synthetic data
- 32k context window with a 131k-vocabulary Tekken tokenizer
What it is
Mistral Small 3 is a 24-billion-parameter language model built by Mistral AI, a French company headquartered in Paris. It arrived in January 2025 as two downloadable checkpoints: a pretrained base and an instruction-tuned version, both released under the Apache 2.0 licence. Mistral AI describes it as aimed at the bulk of everyday generative AI work, where strong language and instruction following matter more than heavyweight reasoning. The company claims it matches models three times its size while running far faster on the same hardware, and that it is a practical open substitute for smaller proprietary models. Quantised, it fits on one RTX 4090 graphics card or a MacBook with 32GB of memory.
How it came about
Mistral AI introduced the model as a latency-optimised addition to its open line-up, renewing a stated commitment to Apache 2.0 for general-purpose models as it moves away from its own MRL licence. Both checkpoints were published at once, and the company said the model sits earlier in its production pipeline than reasoning-focused releases such as DeepSeek R1, since it was trained without reinforcement learning or synthetic data. It was made available on Mistral's la Plateforme and, from day one, through partners including Hugging Face, Ollama, Kaggle, Together AI and Fireworks AI, with NVIDIA NIM, Amazon SageMaker, Groq, Databricks and Snowflake announced as forthcoming.
Inside the model
The company says the model uses far fewer layers than rivals of similar size, which cuts the time taken for each forward pass and drives latency down to around 150 tokens per second. It was trained without reinforcement learning and without synthetic data, leaving it as a base on which others can build reasoning ability. Text is handled by a Tekken tokenizer with a 131k vocabulary, and the context window is 32k tokens. The instruction-tuned checkpoint is a fine-tune of the matching base model. The model supports system prompts, native function calling and JSON output, and the company recommends a low sampling temperature, around 0.15, when serving it.
What it can do
Mistral AI reports over 81% accuracy on MMLU. In its own evaluation pipeline the instruction-tuned model scored 0.663 on MMLU-Pro, 0.453 on GPQA, 0.848 on HumanEval, 0.706 on MATH, 8.35 on MT-Bench, 52.27 on WildBench, 0.873 on Arena-Hard and 0.829 on IFEval. The company says it performs competitively against open models several times larger and against GPT-4o-mini across code, maths, knowledge and instruction following. It also ran side-by-side human comparisons with an external vendor on over a thousand proprietary prompts, acknowledging that human judgements can diverge sharply from public benchmarks. Suggested uses include fast conversational agents, function calling, domain fine-tuning and private local inference.
How to use it
Both the pretrained and instruction-tuned checkpoints are downloadable under Apache 2.0, free to modify and use commercially. The model is served on Mistral's la Plateforme as mistral-small-latest or mistral-small-2501, and is distributed through Hugging Face, Ollama, Kaggle, Together AI and Fireworks AI. It runs with vLLM and transformers, and the company points to mistral-common as the reference for its chat template. Serving it at full precision needs roughly 55GB of GPU memory in bf16 or fp16, though quantised builds run on a single RTX 4090 or a 32GB MacBook.
Reception and debate
Mistral AI's own side-by-side testing, run with an external vendor on more than a thousand anonymised coding and generalist prompts, had evaluators prefer its output over Gemma-2-27B and Qwen2.5-32B in a majority of cases, while results against Llama-3.3-70B and GPT-4o-mini were closer, with those models preferred more often than not. The company notes that human-judgement results can differ starkly from public benchmarks and says it took care to verify the comparisons were fair. No independent reception is described here.
Why it matters
Mistral AI presents the release as a complement to large open reasoning models such as DeepSeek's, and as a strong base for building reasoning capabilities on top. The company says customers are already testing it in financial services for fraud detection, in healthcare for customer triage, and in robotics, automotive and manufacturing for on-device command and control, alongside customer service and sentiment analysis. It also flagged further small and large Mistral models with stronger reasoning as coming soon after release.
Questions people ask about Mistral Small 3
What is Mistral Small 3?
It is a 24-billion-parameter language model from Mistral AI, released in January 2025 in both a pretrained and an instruction-tuned form. The company targets it at everyday generative AI tasks where speed matters, and it is published under the Apache 2.0 licence.
How do I run it?
Weights are free to download and deploy locally. Quantised, the model fits on a single RTX 4090 or a MacBook with 32GB of RAM; at full precision, serving needs about 55GB of GPU memory. It also runs through Mistral's la Plateforme and partner platforms.
What can it do?
It handles text generation in dozens of languages, follows system prompts, and supports native function calling and JSON output for agentic workflows. Mistral AI suggests fast conversational assistants, low-latency tool use, and fine-tuning into domain specialists such as legal or medical helpers.
How does it compare to bigger models?
Mistral AI says it is on par with Llama 3.3 70B instruct while running more than three times faster on the same hardware, and competitive with GPT-4o-mini. Its own human comparisons favoured it over Gemma-2-27B and Qwen2.5-32B more often than not.
Was it trained with reinforcement learning?
No. Mistral AI states the model used neither reinforcement learning nor synthetic data, which places it earlier in its production pipeline than reasoning-focused releases like DeepSeek R1. The company positions it as a base for others to build reasoning ability on.
Good reading and sources
There is no English Wikipedia article on Mistral Small 3 yet, so the facts on this page come from the sources listed above: mistral.ai/news/mistral-small-3, huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.