The last and strongest edition of Mistral's separate reasoning family: 24 billion parameters, vision, a readable chain of thought and an Apache 2.0 licence — withdrawn from the vendor's API in July 2026.
Magistral Small 1.2 is the September 2025 revision of the small reasoning model that Mistral AI built on top of Mistral Small 3.2. It was trained by supervised fine-tuning on reasoning traces from the larger Magistral Medium and then reinforcement learning on top, which is the standard recipe for turning a general chat model into one that thinks step by step before answering.
Two things separate this revision from its two predecessors. It gained a vision encoder, so the chain of reasoning can run over an image and not only over text. And the reasoning trace was wrapped in dedicated [THINK] and [/THINK] tokens, which sounds like an implementation detail but decides whether an application can reliably separate the model's deliberation from its answer — earlier versions marked the trace with ordinary text that the model could also produce by accident.
The vendor's own benchmark table shows the practical gain: on the AIME 2024 mathematics set the score rose from 70.52% in version 1.1 to 86.14% here, and on LiveCodeBench from 59.17% to 70.88%. That put the 24-billion-parameter open model within a few points of Magistral Medium 1.1, the closed model of the previous generation.
The profile is filed as historical for a reason that matters more than the model itself. Mistral deprecated Magistral Small 1.2 in its API on 30 April 2026 and switched it off on 31 July 2026, pointing users to Mistral Small 4. Every other member of the family — Small 1.0, Small 1.1, Medium 1.0, 1.1 and 1.2 — is in the same retirement table. The lab no longer sells a separate reasoning line: reasoning is now a variant of the mainline models, published as Ministral 3 Reasoning checkpoints and folded into Mistral Small 4. The Apache 2.0 weights, however, remain downloadable and usable, so the model is retired as a service, not as software.