Can RTX 4090 24GB run Mistral Small 3.2 24B?
Measured results for Mistral Small 3.2 24B (24.01B dense) on RTX 4090 24GB, plus what fits and the commands to run it.
Short answer
Yes. Mistral Small 3.2 24B Q4_K_M fits fully on RTX 4090 24GB at 16k context, using 17.0 of 23.0 GB usable, and generates about 61.6 tokens/s in our own run. As the context fills it slows down: 51.8 tok/s with 16k tokens already in context (84% of the empty-context speed) and 45.8 tok/s at 32k (74%), measured. For Q6_K, our bandwidth-only estimate was ~29 tok/s against a measured median of ~46 (37% too pessimistic).
Measured generation speed by quantization
| Quant | Generation tok/s | Range | Prompt tok/s | Runs | Origin |
|---|---|---|---|---|---|
| Q4 class (median) | 61.6 | 61.6–61.6 | 3303 | 1 | includes our runs |
| Q4_K_M (our run, depth 0) | 61.6 | 3 repeats | 3303 | 3 | own measurement |
Generation = tg128 (128 generated tokens, short context). Public rows are medians of community llama.cpp results (sources); our own runs use llama-bench b11658 on a rented GPU.
Speed vs context depth measured
Our own llama-bench runs on GeForce RTX 4090 (flash attention on, 3 repeats): generation and prompt speed with the given number of tokens already in the context.
| Quant | Context depth | Generation tok/s | vs depth 0 | Prompt tok/s |
|---|---|---|---|---|
| Q4_K_M | 0 | 61.6 | 100% | 3303 |
| Q4_K_M | 4k | 58.7 | 95% | 3404 |
| Q4_K_M | 16k | 51.8 | 84% | 2791 |
| Q4_K_M | 32k | 45.8 | 74% | 2557 |
Alternatives
Less memory that still fits Mistral Small 3.2 24B
Smaller memory usually means a cheaper card or Mac (we do not list prices).
- Mac mini M6 32GB: Q4_K_M, 22 GB usable, ~7 tok/s estimated
- MacBook Air M5 32GB: Q4_K_M, 22 GB usable, ~6 tok/s scaled from measured
Other models on RTX 4090 24GB, similar size
- LFM2.5 24B-A2B (hardware page): Q6_K, fits fully, ~265 tok/s (scaled from measured)
- Gemma 4 26B-A4B: UD-Q5_K_M, fits fully, ~135 tok/s (scaled from measured)
- gpt-oss 20B: Q8_0, fits fully, ~258 tok/s (measured)
- Qwen3.6 27B: Q5_K_M, fits fully, ~42 tok/s (scaled from measured)
Memory and context limits
Q6_K weights are 19.4 GB; the 16k f16 KV cache adds 2.68 GB. Largest context that fits in 23.0 GB usable:
| KV cache type | Max context |
|---|---|
| f16 | 16k |
| q8_0 | 32k |
| q4_0 | 64k |
Native limit 128k. Full KV cache table for Mistral Small 3.2 24B.
Copy-paste commands
The advisor's pick for this hardware: Q6_K, 16k context.
llama-server -hf unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:Q6_K -c 16384 -fa on --jinja -ngl 99
OLLAMA_CONTEXT_LENGTH=16384 ollama serve ollama run hf.co/unsloth/Mistral-Small-3.2-24B-Instruct-2506-GGUF:Q6_K
FAQ
How fast is Mistral Small 3.2 24B on RTX 4090 24GB?
About 61.6 tokens/s for generation (Q4_K_M on llama.cpp CUDA, 128-token generation), 3303 tokens/s for prompt processing. Speed drops as the context fills; see the depth table.
How much memory does Mistral Small 3.2 24B need on RTX 4090 24GB?
The advisor's pick, Q6_K, is 19.4 GB of weights plus 2.68 GB of KV cache at 16k context (22.0 GB), against 23.0 GB usable. The measured Q4_K_M run needs 17.0 GB.
Is the Mistral Small 3.2 24B speed on RTX 4090 24GB measured or estimated?
Measured: our own llama-bench run on a rented GeForce RTX 4090, including runs with 4k to 32k tokens already in context.
What can I use instead of RTX 4090 24GB for Mistral Small 3.2 24B?
With less memory, Mac mini M6 32GB (~7 tok/s), MacBook Air M5 32GB (~6 tok/s) also fit it fully.
What is the maximum context for Mistral Small 3.2 24B on RTX 4090 24GB?
With an f16 KV cache about 16k tokens, 32k with q8_0 and 64k with q4_0, bounded by the model's own limit.
Related
All models on RTX 4090 24GB · Mistral Small 3.2 24B on all hardware · VRAM tier · all measured pages · leaderboard
Compare: RTX 3090 24GB vs RTX 4090 24GB · RTX 4090 24GB vs RTX 5090 32GB · RTX 4090 24GB vs AMD RX 7900 XTX 24GB · RTX 5080 16GB vs RTX 4090 24GB