Mistral Small 3 Release

Mistral Small 3 is a 24-billion-parameter model released by Mistral AI on January 30, 2025 under the Apache 2.0 license. It is a latency-optimized, pre-trained and instruction-tuned model positioned as an open replacement for proprietary models such as GPT-4o-mini. The release includes both a pretrained and an instruction-tuned checkpoint.

Mistral Small 3 achieves over 81% accuracy on MMLU and a throughput of 150 tokens per second, which Mistral describes as making it "currently the most efficient model of its category." It performs comparably to Llama 3.3 70B Instruct while running more than 3x faster on the same hardware. The model supports text generation, instruction following, code, math, function calling, and agentic workflows, and can run locally on a single RTX 4090 or a MacBook with 32 GB RAM when quantized.

Human evaluations were conducted via an external third-party vendor using side-by-side comparisons on over 1,000 proprietary coding and generalist prompts, with evaluators selecting preferred responses from anonymized outputs. Benchmark scores were obtained through a single internal evaluation pipeline, and Mistral notes that numbers may vary from previously reported figures for competing models due to pipeline differences. Judge-based evaluations (Wildbench, Arena Hard, MTBench) used gpt-4o-2024-05-13 as the judge.

The card does not discuss safety-specific red-teaming, catastrophic-risk evaluations, or CBRN/cyber/autonomy assessments.

The card does not disclose deployed safety classifiers, refusal training, content filters, or any ASL/FSF tier designation.

The model is available on Mistral's la Plateforme under the API identifiers mistral-small-latest and mistral-small-2501, and weights are freely downloadable and modifiable under Apache 2.0. At launch it is also available on Hugging Face, Ollama, Kaggle, Together AI, Fireworks AI, and IBM Watson X, with NVIDIA NIM, Amazon SageMaker, Groq, Databricks, and Snowflake listed as coming soon. Mistral states the model is free to use, modify, and deploy locally with no usage restrictions under Apache 2.0.

Mistral states that Mistral Small 3 is "neither trained with RL nor synthetic data," placing it "earlier in the model production pipeline" than models such as DeepSeek R1, and positions it as a base for further reasoning capability development rather than a finished reasoning model. No additional limitations are disclosed in the document.

This is the initial release of the Mistral Small 3 model series. The release also marks Mistral's stated commitment to "progressively move away from MRL-licensed models" in favor of Apache 2.0 for general-purpose models. No prior-version changelog is included.