Mistral Small 3.1 24B

Mistral Small 3.1-24B Instruct, an upgraded variant of Mistral Small 3 (2501), brings together 24 billion parameters, native image-text understanding, and a 128k token context window to deliver state-of-the-art performance across reasoning (MMLU 80.62%, GPQA Diamond 45.96%), coding (HumanEval 88.41%, MBPP 74.71%), math (MATH 69.3%), and multilingual benchmarks (71.18% average across 24 languages), while also dominating vision tasks like MathVista (68.91%), DocVQA (94.08%), and AI2D (93.72%), outperforming similarly sized proprietary models like Gemma 3 27B and GPT-4o Mini in most categories, all under an Apache 2.0 license and runnable on a single RTX 4090 or 32GB MacBook—making it uniquely positioned for fast, private, and open deployment in applications ranging from conversational agents and function calling to document analysis and edge-device multimodal inference.