Question 1

Can I run Devstral Small 2 24B on a Apple M3 Ultra?

Accepted Answer

Yes — Apple M3 Ultra runs Devstral Small 2 24B fully on GPU at the FP16 quantization. Full GPU inference — 192 GB VRAM meets the 62.4 GB recommendation.

Question 2

What quantization of Devstral Small 2 24B should I use on a Apple M3 Ultra?

Accepted Answer

For 192 GB VRAM on the Apple M3 Ultra, the FP16 variant is the best fit. Estimated ~16.7 tokens/sec on the FP16 quantization.

Question 3

How fast does Devstral Small 2 24B run on Apple M3 Ultra?

Accepted Answer

Roughly 16.7 tokens/sec for FP16. Real speed depends on context length, backend (Ollama, llama.cpp, LM Studio), and KV cache size.

Question 4

What if Apple M3 Ultra is not enough for Devstral Small 2 24B?

Accepted Answer

Try a smaller quantization variant or run on CPU + RAM with the trade-off of slower generation.

Quantization	File Size	Min VRAM	Rec VRAM	Context	Verdict	Estimated tok/s
Q4_K_M	12 GB	13.8 GB	15.6 GB	8K / 8K	Full GPU	~53.3
Q5_K_M	15 GB	17.3 GB	19.5 GB	8K / 8K	Full GPU	~46.4
Q8_0	24 GB	27.6 GB	31.2 GB	8K / 8K	Full GPU	~31.7
FP16Best fit	48 GB	55.2 GB	62.4 GB	8K / 8K	Full GPU	~16.7

Can I Run Devstral Small 2 24B on Apple M3 Ultra?

Share this matchup

Every Devstral Small 2 24B quantization on Apple M3 Ultra