Builds / Apple M5 Ultra (unified, Mac Studio)
Mac Studio M5 Ultra 256 GB: Qwen3.6 35B-A3B at 157 tok/s on MLX 4-bit, 142 NVFP4, 118 MXFP8, 134-136 on llama.cpp GGUF (URE)
Last updated:
- GPU(s)
- Apple M5 Ultra (unified, Mac Studio)
- VRAM / unified memory (GB)
- 256
- CPU
- Apple M5 Ultra
- Motherboard
- Mac Studio
Amazon · eBay (used) · Newegg - RAM
- 256 GB unified (about 1.1 TB/s measured)
- PSU
- not stated
- Other
- Mac Studio M5 Ultra, 256 GB; MLX / mlx_lm, vLLM 0.32 + Metal plugin, llama.cpp GGUF
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.6 35B-A3B | MLX affine 4-bit | MLX | 157 | 8,700 @ 8,192-token prompt est. TTFT @8k ≈ 0.94 s (calculated from measured prefill) | context: single stream📏 Single-stream decode, 6.4 ms/token; 8K-prompt prefill about 7.8k-8.7k tok/s; a later clean session measured 163 |
| Qwen3.6 35B-A3B | MLX NVFP4 | MLX | 142 | not stated | context: single stream📏 Single-stream decode; 151 in a later clean session; 8K-prompt prefill about 8.6k tok/s |
| Qwen3.6 35B-A3B | MLX MXFP8 | MLX | 118 | not stated | context: single stream📏 Single-stream decode; 8K-prompt prefill about 7.2k tok/s |
| Qwen3.6 35B-A3B | GGUF Unsloth UD-Q4_K_XL | llama.cpp | 134 | not stated | context: single stream📏 Single-stream decode; MXFP4_MOE GGUF 136; 8K-prompt prefill about 4.0k tok/s |
Prompt processing = prefill speed (how fast the prompt is read) as reported by the source, with the prompt length it was measured at; blank = the source does not report it. est. TTFT @8k is calculated from the measured prefill speed (8,192 ÷ prefill tok/s), shown only when prefill was measured on a prompt of 512+ tokens. It is not a measurement and ignores load time and caching. measured TTFT is the source's own time-to-first-token. Why prefill matters for coding assistants.
Owner-run test by Stefano Schotten (URE, 5 Nov 2026), same model and evening as an RTX PRO 6000 Max-Q (a separate RigBase build). On a real multi-turn RAG workload (about 11,300-token prompts, 200-token answers) the Mac's best engine delivered 51% of the card's throughput for one user (78% with the author's mlx-lm MTP patch) and about 30% under load. Pure batched decode (bare MLX kernel): 151 tok/s at batch 1, 1,278 aggregate at batch 64. The prefill figure stored here is the upper end of the stated range.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Quick answers: Can Apple M5 Ultra run Qwen3.6 35B-A3B?