Magistral Small 1.2

Magistral Small 1.2 (API id magistral-small-2509) was released by Mistral AI on September 18, 2025, as an update to its open reasoning line. The 24B dense model adds a visual encoder for image analysis, extends context to 128K tokens, and introduces a [THINK] token that encapsulates the inference process.

It ships under the Apache 2.0 license with inference templates for vLLM, Transformers, and llama.cpp, plus a GGUF quantization that fits on a single RTX 4090 or a 32GB MacBook. Mistral positions it as a small reasoning model competitive with much larger ones on math and logic benchmarks.