Coding with AI / Mac Studio M3 Ultra
Mac Studio M3 Ultra
Memory: 96 / 256 / 512 GB.
Memory bandwidth: 819 GB/s.
Generation speed is bound by the chip's bandwidth; memory decides which models fit.
smart and fastQwen3.6-35B-A3B51.5%* on Terminal-Bench 2~122 tok/s · estimated How to run
Qwen3.6-35B-A3B smart
-
Score: 51.5% (Terminal-Bench 2),
source
-
Speed: 122 tok/s (estimate: bandwidth-bound approximation)
- Conditions: engine llama.cpp, quant Q4_K_M, context 128
- Min memory: 24 GB
- Command:
llama-server -m qwen3.6-35b-a3b-Q4_K_M.gguf
GLM-5.2 ceiling
-
Score: 82.7% (Terminal-Bench 2),
source
-
Speed: 9 tok/s (estimate: bandwidth-bound approximation; ~40B active params per community reports, vendor does not publish them)
- Conditions: engine llama.cpp, quant UD-Q4_K_M, context 128
- Min memory: 512 GB
- Command:
llama-server -m glm-5.2-UD-Q4_K_M.gguf
Every number above is either a measurement with a source or a labeled estimate with a method.
Unlabeled estimates do not exist on this site.
Ceiling timeline
The strongest model fitting Mac Studio M3 Ultra (512 GB) at its release date, by SWE-bench Verified.
- 2025-07-22 Qwen3-Coder-480B 69.6%
- 2025-12-09 Devstral 2 72.2%
- 2025-12-22 GLM-4.7 73.8%