Mistral AI releases Mistral Small 3 under Apache 2.0

Background

Mistral AI had been publishing its general-purpose models under a research licence, a set of terms that limits what licensees may do with the weights. Alongside Mistral Small 3, the company said it would move its general-purpose models from that research licence to Apache 2.0 ^[1]. Apache 2.0 is a permissive open-source licence: it lets anyone use, modify and redistribute the covered work, including commercially, as long as the licence text and attribution notices are preserved. For model weights, the practical difference is who can build on them.

What was released

On 30 January 2025, Mistral AI released Mistral Small 3, a model with 24 billion parameters, under Apache 2.0 ^[1]. Parameters are the numeric weights a neural network learns during training; a model of this size sits below the largest systems and above the smallest ones, and the count largely determines what hardware is needed to run it.

How it performs

The model scored above 81% on MMLU ^[1], a multiple-choice benchmark that tests knowledge across a wide range of academic subjects, from mathematics and history to law and medicine. Mistral reported that it ran at roughly 150 tokens per second ^[1]. A token is the unit of text a model reads and writes, roughly a short word or a fragment of one, so this figure describes how fast text is generated. The company also said that Mistral Small 3 matched Llama 3.3 70B Instruct while running more than three times faster ^[1].

How it was trained

Mistral said the model was trained neither with reinforcement learning nor with synthetic data ^[1]. Reinforcement learning trains a model by rewarding outputs that score well against some objective, and synthetic data is training text produced by another model rather than written by people. Stating that neither was used distinguishes this model from pipelines that depend on reward-driven post-training or on model-generated corpora.

Running it locally

When quantized, Mistral Small 3 can run privately on a single RTX 4090 or on a MacBook with 32 GB of RAM ^[1]. Quantization stores the weights at lower numerical precision, trading a little accuracy for a much smaller memory footprint, which is what allows a model of this size to fit on consumer hardware instead of a data-centre GPU.

Why it matters

The licence and the size work together. Apache 2.0 permits modification and commercial use ^[1], and a 24-billion-parameter model that fits, once quantized, on one consumer graphics card or a laptop ^[1] can be deployed on hardware an organisation already owns. Running weights locally keeps prompts and data inside a private network rather than sending them to an external service. Mistral's comparison point was Llama 3.3 70B Instruct, a much larger model, and the company emphasised speed rather than only quality ^[1].