Muse Spark runs on a single Mac Studio M3 Ultra at 18 t/s decode
challis88ocarina · reddit · 2026-10-18
A Reddit user reports running Muse Spark locally on a single Mac Studio (M3 Ultra, 256GB unified memory):
- Decode speed hit 18 t/s, with coherent text output and solid tool calls (text-only)
- A slight regression appeared after cherry-picking code from other DeepSeek V4 (ds4) work trees; the author is fixing a prefill droop and will push code once 18 t/s is restored
- Weights are gated on Hugging Face, so the author is weighing uploading weights vs. publishing a reproduction recipe
The attached log shows the Metal backend mapping 205 GiB of resident model memory, 85 t/s average prefill over a 21,346-token prompt, 262k context, and decoding around 13.4 t/s.
More from Infra
- Modal triples valuation in a week while data center owners raise nothing — YvesMulkers · 2026-10-18
- Reddit asks: is 2-bit or 3-bit quantization actually usable for daily LLM work? — Queasy_Back2588 · 2026-10-18
- Samsung and TSMC reportedly team up to produce Nvidia's next-gen HBM chips — Polymarket · 2026-10-18
- Safari Caps Incognito and Low-Power Mode at 30fps — Found via AI-Assisted Bench Bisect — oykun · 2026-10-18
- Running DeepSeek V4.1-Flash (318GB) on a single modded 4090 at 20+ tok/s — lxfater · 2026-10-18
- Hana Securities: YMTC likely to IPO this December, aiming to raise $33B — basedjensen · 2026-10-18