Can I run it?
Can I run gemma 4 31B it on Apple M5 Max (40-core GPU)?
Tight fit
Tight fit fully in unified memory on your Apple M5 Max (40-core GPU) at UD-Q8_K_XL, 8K context, about 7.8–11 tok/s. Weights 32.6 GB plus 2.03 GB of KV cache — 35.3 GB of 36.0 GB usable.
Best quantUD-Q8_K_XL
Working context8K
Max that fits8K
Speed7.8–11 tok/s
Weights32.6 GB
KV cache2.03 GB
Total35.3 GB
Usable36.0 GB
gemma 4 31B it details · Apple M5 Max (40-core GPU) page · How this estimate is made
Run commands
llama.cpp
llama-server -hf unsloth/gemma-4-31B-it-GGUF:UD-Q8_K_XL -c 8192 -ngl 60
Ollama
ollama run gemma
LM Studio
lms get unsloth/gemma-4-31B-it-GGUF # Then load quant UD-Q8_K_XL
Assumptions
- Estimate, not a benchmark. Decode speed uses a memory-bandwidth model at 55–75% of peak.
- KV cache is FP16. Engine overhead assumes llama.cpp (0.3 GB + 1% of the weights). Ollama is closer to 0.9 GB + 3%, LM Studio closer to 1 GB + 3%.
- UD-Q8_K_XL size is the real GGUF file size from Hugging Face.
- Unified memory uses 75% for the GPU by default, and always leaves about 5 GB for the OS.
- Speed is quoted at 8K context, not at the maximum context that still fits.
Larger models that also fit
- Can I run Qwen3.6 35B A3B on Apple M5 Max (40-core GPU)? — Tight fit, Q8_0, 100–137 tok/s
- Can I run Qwen3.8 35B A3B Distill on Apple M5 Max (40-core GPU)? — Runs, Q6_K, 128–175 tok/s
- Can I run Qwen3.5 35B A3B on Apple M5 Max (40-core GPU)? — Tight fit, Q8_0, 100–137 tok/s
Updated 2026-10-24. Profile on this page is the device default.