Mistral Small 3.2

What it is

Mistral Small 3.2 is Mistral AI's compact open-weights model: fast and cheap to run with dependable quality, as the note on this page puts it. The Small line is the part of Mistral's range that actually ships weights, and it has been a staple of self-hosted deployments since the first Mistral 7B — small enough for modest hardware, serious enough to be useful.

The 3.2 suffix implies a point release in a long series. Check Mistral's channels for what changed against 3.1 and whether the licence for this specific release differs from its predecessors.

How the composite is built

Overall 79 weights Coding 24%, Terminal 20%, Reasoning 20%, Tool use 15%, Context 13%, Speed 8%. Capability axes range from 74 to 80; Speed 88 is nearly twenty points clear of the rest of the profile and does most of the work in the composite.

Coding 79

Level with Phi-5 and Grok 5 Mini. Boilerplate, tests and refactors with review, rather than novel design.

Terminal 74

Level with Gemma 3 27B and Yi-2 Large, and the model's weakest axis. Short, checked shell commands only.

Reasoning 78

One behind Grok 5 Mini's 79 and six behind Phi-5's 84 — the small tier's reasoning outlier. Task-scale planning, not deep analysis.

Tool use 79

One above Grok 5 Mini's 78 and one below Amazon Nova 2 Pro's 80. Dependable in short chains with schema validation.

Context 80

Just below the 82 shared by several open models here. File-scale work; use retrieval for anything larger.

Speed 88

Level with Gemini 3.6 Flash's 88 and behind only the 90-plus small tiers. On hosted rates this is the quickest way to buy Mistral-quality output.

Where it fits

High-volume, low-stakes work at the cheapest dependable rate on this board: classification, extraction, summarisation, drafting support. The open weights make it equally at home on a single GPU where the data cannot leave the building — the same niche Gemma 3 27B occupies.

Limits

• Terminal 74 is the real constraint on any agentic use. This is a batch model, not an operator.
• Value is Unrated because no Artificial Analysis index is published for it — at $0.13/M blended the price is the lowest of any scored model here, so do not read Unrated as expensive.
• Quality drops visibly against the Large tier — four to six points on every axis. Route hard work up the stack.
• Successor releases are already on the board unscored — Mistral Small 4 at $0.15/$0.60. Treat 3.2 as a point-in-time read.

Price and access

$0.09 per million input tokens and $0.25 per million output, blending to $0.13/M — the cheapest blended rate among the models scored here. Listed on OpenRouter alongside Mistral's own platform, with weights published for self-hosting. Last scored 15 Sep 2026.

Alternatives on this board

• Gemma 3 27B — 79 at $0.08/$0.45 with open weights from Google. The near-identical profile and price from another vendor.
• Xiaomi MiMo 2.5 Pro — 82 at $0.43/$0.87. Three points up for four times the blended rate.
• Mistral Large 3 — 83 at $0.50/$1.50, the same vendor's balanced tier.
• Phi-5 — 78 with Reasoning 84 and Speed 90, if the small model needs to think rather than transact.

Sources