Mistral Small 3

What it is

Mistral Small 3 is a 24-billion-parameter language model released by Mistral AI on 30 January 2025 under the Apache 2.0 licence, with both pretrained and instruction-tuned checkpoints published.^[1] Mistral built it as a latency-optimised model: the goal was not the largest possible model, but one that answers quickly enough to sit inside interactive applications while still handling general-purpose tasks.^[1]

MMLU, the benchmark most often quoted for this class of model, is a large set of multiple-choice questions spanning many academic subjects. It measures breadth of knowledge rather than any single skill.

Background

The name had been used before for a different kind of product. Mistral's earlier Mistral Small API models were proprietary: mistral-small-2402 appeared in February 2024 and mistral-small-2409 in September 2024, both served only through Mistral's own API rather than as downloadable weights.^[2][4] Mistral Small 3 broke with that pattern by shipping open checkpoints under a permissive licence.^[1]

Performance

Mistral reports that the model scored over 81% on MMLU while generating at about 150 tokens per second.^[1] The company also said the instruction-tuned version matches Llama 3.3 70B Instruct in quality while running more than three times faster on the same hardware, and that it competes with GPT-4o mini.^[1] Those comparisons are Mistral's own; the point they are making is about the ratio of capability to speed and memory rather than about raw scale.

How it was trained

Mistral states that the model was trained without reinforcement learning and without synthetic data, and positions it as a base on which others can build reasoning models.^[1] Reinforcement learning and machine-generated training data are common ways to lift benchmark scores; training without them keeps the model closer to its pretraining data and, in Mistral's framing, makes it a cleaner starting point for further post-training by whoever adopts it.^[1]

Where it runs

A central claim of the release is that the model is small enough to be practical outside a data centre. Quantized, it runs on a single RTX 4090 graphics card or on a 32GB MacBook.^[1] Quantization here simply means storing the model's weights at lower precision so it takes less memory, at some cost in numerical fidelity.

At launch the model was available on La Plateforme, Hugging Face, Ollama, Kaggle, Together AI, Fireworks AI and IBM watsonx.^[1] Mistral also said it was coming soon to NVIDIA NIM, Amazon SageMaker, Groq, Databricks and Snowflake.^[1]

A licensing shift

Alongside the release, Mistral said it was moving its general-purpose models to Apache 2.0, away from the Mistral Research License.^[1] Apache 2.0 permits commercial use and modification with few conditions, whereas a research licence typically restricts how weights may be used in products. For developers, the change matters as much as the model itself: it determines whether a model can be embedded in a shipped product without a separate agreement.

Retirement

Mistral Small 3 had a short life in Mistral's catalogue. It was deprecated on 6 November 2025 and retired on 30 November 2025, replaced by Mistral Small 4.^[3] The weights released under Apache 2.0 remain downloadable, but the model is no longer the current entry in its size class.