Builds / Apple M5 (unified)
MacBook Pro M5 32 GB: Qwen3.8 Flash-Next (125B-A6B) IQ3_XXS at 10.2 tok/s and Qwen3.6 35B-A3B Q6 at 16.1, experts streamed from SSD (MoEShift)
Last updated:
- GPU(s)
- Apple M5 (unified)
- VRAM / unified memory (GB)
- 32
- CPU
- Apple M5
- Motherboard
- MacBook Pro
Amazon · eBay (used) · Newegg - RAM
- 32 GB unified
- PSU
- not stated
- Other
- macOS 27.2, MoEShift built from source (2b1b3e8), default config (no manual tuning); GGUF and extracted experts on the internal SSD; on mains power, Low Power Mode off
- 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 | ISTA-DASLab GSQ-RCO IQ3_XXS (~76 GB, 2 shards), experts streamed from the internal SSD | MoEShift source build 2b1b3e8 (llama.cpp-based, Metal) | 10.15 | 32.33 @ 72-token prompt | context: 72-token prompt, 512 generated📏 moeshift-community-v1 protocol: median of 3 measured runs after 1 warm-up, 512 output tokens; system idle (issue #6) |
| Qwen3.8 Flash-Next | ISTA-DASLab GSQ-RCO IQ3_XXS (~76 GB, 2 shards), experts streamed from the internal SSD | MoEShift source build 2b1b3e8 (llama.cpp-based, Metal) | 9.21 | 201 @ 6K-token prompt est. TTFT @8k ≈ 40.7 s (calculated from measured prefill) | context: 6,060-token prompt, 512 generated📏 moeshift-community-v1 protocol: median of 3 measured runs after 1 warm-up, 512 output tokens; system idle (issue #6) |
| Qwen3.6 35B-A3B | Unsloth UD-Q6_K_XL (single file), experts streamed from the internal SSD | MoEShift source build 2b1b3e8 (llama.cpp-based, Metal) | 16.07 | 40.71 @ 72-token prompt | context: 72-token prompt, 512 generated (64K ctx configured)📏 moeshift-community-v1 protocol: median of 3 measured runs after 1 warm-up, 512 output tokens; normal work session (IDE, video playing); peak RSS ~7.7 GB, no swap (issue #7) |
| Qwen3.6 35B-A3B | Unsloth UD-Q6_K_XL (single file), experts streamed from the internal SSD | MoEShift source build 2b1b3e8 (llama.cpp-based, Metal) | 14.94 | 474 @ 6K-token prompt est. TTFT @8k ≈ 17.3 s (calculated from measured prefill) | context: 6,060-token prompt, 512 generated📏 moeshift-community-v1 protocol: median of 3 measured runs after 1 warm-up, 512 output tokens; same session (issue #7) |
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.
Two community reports by the same tester on the MoEShift GitHub tracker (issues #6 and #7, filed 10-11 Nov 2026). MoEShift is an open-source llama.cpp-based engine for Apple Silicon that keeps shared weights in memory and streams Mixture-of-Experts weights from the SSD as needed, so the ~76 GB Flash-Next file runs on a 32 GB base-M5 laptop. Speed depends on SSD speed and expert cache hits. IQ3_XXS is below Q4. The Qwen3.6 35B-A3B Q6 file (~30 GB) would nearly fill 32 GB if loaded normally; with MoEShift peak RSS stayed ~7.7 GB.
https://github.com/davide-troiani/moeshift/issues/7
https://github.com/davide-troiani/moeshift
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.