Mistral Small 3.2 24B: The Point Release Agents Needed

Point releases rarely get attention, and most do not deserve it. Mistral Small 3.2 is an exception worth understanding, because what it fixed is not raw intelligence but the failure modes that make a model unusable inside a loop: ignoring precise instructions, getting stuck repeating itself, and emitting malformed tool calls.

The model is published as Mistral-Small-3.2-24B-Instruct-2506 — 24B parameters, Apache 2.0, released in June 2025 as a direct update to Mistral Small 3.1. Same size, same image input, different behaviour.

Three fixes, all aimed at reliability

The model card lists exactly three changes from 3.1:

• Instruction following. Better at following precise instructions. Mistral's internal instruction-following accuracy rises from 82.75% to 84.78%.
• Repetition. Mistral describes infinite or repetitive generations as reduced about 2x on challenging prompts; the table's own measure falls from 2.11% to 1.29%.
• Function calling. A more robust calling template.

Everything else is expected to match or slightly improve on 3.1. That is an unusually honest framing for a release note, and the benchmark table bears it out.

A 1–2% repetition rate sounds small until you multiply it. An agent making a hundred calls per task at 2.11% will, on average, hit a runaway generation on about two of them. Each one burns output tokens up to your max_tokens limit and usually derails the step. Cutting that rate by roughly 40% is worth more to an agent than several points on a knowledge benchmark. Detecting agent loops covers catching these before they cost you.

Reading the before-and-after table

The useful thing about the 3.2 model card is that it reports both versions on the same harness:

• WildBench v2 — 55.6% to 65.33%
• Arena Hard v2 — 19.56% to 43.1%
• MMLU-Pro (5-shot CoT) — 66.76% to 69.06%
• HumanEval Plus (pass@5) — 88.99% to 92.90%
• MBPP Plus (pass@5) — 74.63% to 78.33%
• MMLU — 80.62% to 80.50%; GPQA Diamond — 45.96% to 46.13%
• MMMU — 64.00% to 62.50%; MathVista — 68.91% to 67.09%; DocVQA — 94.08% to 94.86%

The pattern is clear. Knowledge and reasoning are flat — MMLU and GPQA barely move, which is what you expect when the base model has not changed. Chat-quality and coding scores rise, which is what better post-training produces. Vision is mixed, with small losses on MMMU and MathVista.

The large Arena Hard jump deserves some caution. It is a judge-scored benchmark, and post-training that improves formatting and tone moves judge scores more than it moves correctness. A doubled preference score and a flat MMLU can both be true at once.

Serving it yourself

The weights are Apache 2.0, and the model card gives a concrete recipe. It recommends vLLM 0.9.1 or later and serves with the Mistral-specific tokenizer, config and load formats, plus --tool-call-parser mistral and --enable-auto-tool-choice so tool calls come back structured. The example uses --tensor-parallel-size 2.

Running it in BF16 or FP16 needs about 55 GB of GPU memory, according to the card. That rules out one 48 GB card unquantised, but quantised builds bring it within reach of a single workstation GPU, at a quality cost worth measuring on your own prompts.

Two settings are worth copying exactly:

1. Temperature 0.15. Mistral recommends it. Higher values undo some of the reliability gains you chose this model for.
2. The Mistral tool parser. Without it, tool calls arrive as text that your client must parse, and the robustness improvement in 3.2 is largely wasted.

On context length, the 3.2 card does not state a figure; 3.1 was announced at 128K and 3.2 is a drop-in update of it. Hosted listings vary, so check the limit your provider actually enforces rather than assuming. The usable figure is lower than the advertised one anyway, since quality falls off well before the limit.

Availability, and what replaces it

Mistral's documentation lists Small 3.2 as deprecated on 30 April 2026 and retired on 31 July 2026, with Mistral Small 4 named as the replacement. It is no longer on Mistral's own API.

Third parties still host it. At the time of writing it is listed on OpenRouter at about $0.09 per million input tokens and $0.25 per million output, with JSON mode, structured outputs and tool calling listed as supported. That is cheaper than the current listing for the older 3.1.

The replacement is not a like-for-like swap. Mistral Small 4 is a 119B mixture-of-experts model with optional reasoning: cheap to call through an API, but a multi-GPU deployment to self-host. If you run 3.2 on your own hardware precisely because it fits on one or two cards, Small 4 does not drop in.

Who should still use it

• Keep 3.2 if you self-host, need image input and tool calling in a 24B footprint, and a permissive licence matters.
• Move from 3.1 to 3.2 unless your workload is image-heavy and a side-by-side test favours 3.1.
• Move to Small 4 for hosted use, where the larger model costs little more per token and is actively maintained.

Whichever way you go, rerun your own regression set on the change. A point release that fixes repetition can still shift output formatting enough to break a parser. Regression testing model upgrades has the method.