MistralAI released a new Magistral Small 2509
7 hours ago
- #reasoning
- #AI model
- #multimodal
- Magistral Small 1.2 is an efficient reasoning model with 24B parameters, built upon Mistral Small 3.2 and enhanced with reasoning capabilities.
- It can be deployed locally, fitting within a single RTX 4090 or a 32GB RAM MacBook once quantized.
- Updates in version 1.2 include multimodality (vision encoder), better performance, improved tone and persona, finite generation, and special think tokens for reasoning.
- Key features include long reasoning chains, multilingual support (dozens of languages), vision capabilities, Apache 2.0 License, and a 128k context window.
- Benchmark results show significant performance improvements over previous versions.
- Recommended sampling parameters: top_p=0.95, temperature=0.7, max_tokens=131072.
- Basic chat template includes a system prompt with [THINK] and [/THINK] tokens for reasoning traces.
- The model can be used with frameworks like vLLM (recommended), transformers, llama.cpp, and Unsloth GGUF.
- Fine-tuning options include Axolotl and Unsloth.
- Detailed usage examples provided for text, text-image (Pokemon battle, geo trivia, math equations), and transformers integration.