Twenty-four billion parameters, fifteen months on

What it is for

Mistral Small 3.2 is a 24-billion-parameter dense model from Mistral AI with open weights, released on 19 June 2025 and curated here for general and agent work. Unsloth publishes the GGUF build we measure, and Ollama carries it in its library as mistral-small3.2. On 25 September it stood at 198,640 downloads over the last 30 days, up 45 % week on week, with 620 likes on Hugging Face and a trending score of 3.

The move came in one step. Thirty-day downloads sat between 128,000 and 137,000 from 12 to 18 September, then read 185,194 on the evening of the 19th and climbed to 199,655 by the 24th, while the trending score went from 1 and 2 to 3 and 4 over the same days. The Heat Score followed from 27 to 52, and the engine flagged the move as a breakout. The runner-up was its stablemate Devstral Small 2 24B, one point behind with a rise of 24 (32 to 56); then Qwen3 14B with 17, GLM-4.7-Flash with 15, Qwen3 0.6B and LFM2.5 2.6B with 13. gpt-oss 20B and Qwen3.8 27B lost 21 and 36 points inside the window, so their scores did not count.

A dense 24B has no active-parameter trick: all 13.3 GB of its Q4_K_M weights are read for every token. With the context cache computed from the published attention shape and the runtime overhead, the engine puts the total at 15.8 GB at 8K tokens and 20.3 GB at 32K. That is GOOD on the 24 GB cards, RTX 3090, RTX 4090 and RX 7900 XTX, with 5.3 GB to spare at 8K and TIGHT at 32K; EXCELLENT on the RTX 5090 and the Mac Studio M5 Max 36GB at 8K; and comfortable at both context lengths on any 64 GB Mac, the M5 Ultra and the DGX Spark. A 16 GB card is 1.7 GB short at 8K and has to offload part of the weights into system RAM. The model accepts up to 131,072 tokens of context on machines with the memory for it. Renting instead of buying: an RTX 4090 cost $0.59 per hour at the median of verified Vast.ai offers on 25 September at 19:00 UTC.

Watch out

• Not a 16 GB card model: 15.8 GB at 8K against 14.1 GB usable means every 16 GB card, including the new RX 9070 XT and RTX 4070 Ti Super, has to offload. It runs, but not in the way a fit verdict means it.
• The rise starts low and is young: a Heat Score of 27 on 19 September, then one jump of 48,000 downloads in a day and a week of readings since. Real, but one week old.
• The model is fifteen months old. Our readings say people are downloading it again; they do not say why, and the rule does not ask.
• The Mac mini M5 Pro 24GB misses by 0.2 GB: macOS wires only about two thirds of its memory for the GPU, 15.6 GB, against 15.8 GB needed at 8K.
• One measured GGUF, Q4_K_M from unsloth; other quantizations change every number above. No benchmarks and no speed figures: whether it is good at your task is yours to test.
• The rental price is one marketplace at one hour; it moves during the day.