M5 Ultra 256GB early test: Mimo2.6-Flash hits 49 tok/s at 32K context
bakawolf123 · reddit · 2026-09-22
A Reddit user got an M5 Ultra (36/80, 256GB unified memory) and hand-patched mlx-vlm to run Mimo2.6-Flash from original weights, sharing early benchmarks.
- 32K prompt / 64 generated tokens: 1238 tok/s prompt processing, 49.1 tok/s generation average
- 64K prompt / 1024 tokens: 982 tok/s prompt, generation drops to 38.1 tok/s, 90s total
- MTP is enabled but doesn't seem to be working; verdict is "pretty meh, but usable"
- Also posted oMLX numbers for qwen3.8-flash-next fp8 as comparison: 42-44 tok/s output, 177-188GB memory
- No out-of-the-box support exists yet; the raw-weights patch is the author's own work
More from Infra
- Gas-for-AI-datacenters forum co-located with CoRL shows energy demand spillover — carlosdponx · 2026-11-12
- a16z podcast: Mike Volpi on Cisco's 70+ acquisition playbook and the AI buildout — a16z · 2026-11-12
- AMD to pass Intel in server revenue by mid-2027 as share gains 6 points a year — BenBajarin · 2026-11-12
- Study: Reflective Data Center Roofs Actually Cool Daytime Ground by 1.8-2.2°C — SumitGup · 2026-11-12
- Tsinghua's BaGuaLu-2 tames MoE expert-load spikes at runtime, accepted to SC26 — jiqizhixin · 2026-11-12
- antirez backs low-bit DeepSeek 4.1 Flash over Qwen 27B dense on Apple M5 Macs — antirez · 2026-11-12