Builds / Apple M5 Ultra 80-core GPU (unified, Mac Studio)
Mac Studio M5 Ultra 256GB: Qwen3.8 Flash-Next UD-Q4_K_XL at 46.1 tok/s at 32K, 41.7 at 128K (llama.cpp Metal, no MTP)
Last updated:
- GPU(s)
- Apple M5 Ultra 80-core GPU (unified, Mac Studio)
- VRAM / unified memory (GB)
- 256
- CPU
- Apple M5 Ultra
- Motherboard
- not stated
- RAM
- 256 GB unified memory
- PSU
- not stated
- Other
- llama.cpp b11921 (7a2e5ab) Metal; KV f16; MTP off; the model's sparse-attention top-k op ran on the CPU in this build (moved to Metal later by PR #31230)
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Qwen3.8 Flash-Next | Unsloth UD-Q4_K_XL (same 4 split files, SHA256-matched) | llama.cpp b11921 (Metal) | 46.11 | 1,863 @ 32K-token prompt est. TTFT @8k ≈ 4.4 s (calculated from measured prefill) | context: 32K tokens already in the session📏 AI Runner test: Italian novel or C++ source fed into one session in 32K-token steps up to 128K; after each step, the new 32K prefill and the following 256-token generation timed over HTTP streaming; greedy, seed 42, 3 runs per text with a fresh server each time; llama.cpp b11921 (7a2e5ab) Metal, KV f16, batch/ubatch 2048/2048, PLE resident (lazy off), MTP off; measured 27 Oct 2026; novel text; 45.90 on code |
| Qwen3.8 Flash-Next | Unsloth UD-Q4_K_XL (same 4 split files, SHA256-matched) | llama.cpp b11921 (Metal) | 41.74 | 1,540 @ 32K at 128K depth-token prompt est. TTFT @8k ≈ 5.3 s (calculated from measured prefill) | context: 128K tokens already in the session📏 AI Runner test: Italian novel or C++ source fed into one session in 32K-token steps up to 128K; after each step, the new 32K prefill and the following 256-token generation timed over HTTP streaming; greedy, seed 42, 3 runs per text with a fresh server each time; llama.cpp b11921 (7a2e5ab) Metal, KV f16, batch/ubatch 2048/2048, PLE resident (lazy off), MTP off; measured 27 Oct 2026; code text; 41.47 on the novel |
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.
AI Runner (Korean blog), published 27 Oct 2026 and updated 3 Nov 2026. Same file and same cumulative task as the author's earlier M4 Max and Ryzen AI Max+ 395 runs: the M5 Ultra generated 1.37-1.43x faster than the M4 Max and read new text about 3x faster; adding 32K at 128K depth took ~21 s vs ~63 s. Generation dropped only ~10% from 32K to 128K. A download was writing to the same SSD (~13 MB/s) during the run. An Uncensored IQ4_XS file with sparse attention disabled was faster at 32K (51.70) but slower at 128K (38.49); not listed as a headline. Output quality not compared.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.
Copy into a README or blog. Links back to this build. See badge docs (CC BY 4.0 attribution). The SVG does not invent speeds — numbers stay on the build page.
Markdown
[](https://getrigbase.com/builds/mac-studio-m5-ultra-256gb-qwen38-flash-next-llamacpp-ai-runner/)
HTML
<a href="https://getrigbase.com/builds/mac-studio-m5-ultra-256gb-qwen38-flash-next-llamacpp-ai-runner/" rel="noopener"><img src="https://getrigbase.com/badge/mac-studio-m5-ultra-256gb-qwen38-flash-next-llamacpp-ai-runner.svg" alt="Measured on RigBase" height="20"/></a>
Preview: 
· How to use
Quick answers: Can Apple M5 Ultra run Qwen3.8 Flash-Next?