Builds / Apple M5 (unified, MacBook Pro)
MacBook Pro M5 32 GB (LM Studio): Gemma 4 12B at 14 tok/s, Qwen3.8 27B at 5.4 (4-bit)
Last updated:
- GPU(s)
- Apple M5 (unified, MacBook Pro)
- VRAM / unified memory (GB)
- 32
- CPU
- Apple M5
- Motherboard
- MacBook Pro
Amazon · eBay (used) · Newegg - RAM
- 32 GB unified memory
- PSU
- not stated
- Other
- LM Studio; Context Length set to 65,536 for each model (default 8,192 was too small for the agent's ~15k system prompt); 4-bit quants
- Reported total cost
- not stated
- Power
- not stated
Measured speed (generation tokens/sec)
| Model | Quant | Engine | tok/s | Prompt processing (tok/s) | Details |
|---|---|---|---|---|---|
| Gemma 4 12B | 4-bit | LM Studio | 14 | 275 (prompt length not stated) | context: short raw-speed probe (generation vs prefill table)📏 Author's raw-speed table before the coding task: generation 14 tok/s, prefill 275 tok/s; thinking not discussed for this row |
| Qwen3.8 27B | 4-bit (dense) | LM Studio | 5.4 | not stated | context: short raw-speed probe (generation vs prefill table)📏 Author's raw-speed table: generation 5.4 tok/s, prefill 159 tok/s; dense 27B on 32 GB unified memory |
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.
dev.to post of 4 Nov 2026 (local vs cloud coding-agent comparison). Raw-speed table only; the agent's task scores are quality, not tok/s. Not listed: Qwen3 Coder 30B-A3B (52 tok/s / 743 prefill) — not a current RigBase model.
Blank fields = not stated in the source. We never estimate build data; GPU street prices above are separately sourced and dated.