mistral-small-24b-instruct-2501

mistral-small-24b-instruct-2501 (Mistral Small 3) is a 24-billion parameter large language model developed by Mistral AI, specifically optimized for reasoning, low latency, and high-efficiency instruction following. Released under the Apache 2.0 license, it is designed to bridge the gap between compact models and frontier-class performance, achieving results comparable to significantly larger models in tasks such as coding, mathematics, and general knowledge.

The model utilizes a transformer-based architecture with a 32,768-token context window and the Tekken tokenizer, which features a 131k vocabulary size for improved efficiency across multiple languages. It is particularly suited for agentic workflows due to its native capabilities in function calling and structured JSON output generation.

Mistral Small 3 is engineered to be "knowledge-dense," making it viable for local execution on high-end consumer hardware. Its instruction-tuning process focuses on precise adherence to complex directions and maintaining coherence in multi-turn conversations, supporting dozens of languages including major European and Asian languages.