DeepSeek V4 Flash on M2 Ultra: lossless repack to 141GiB, 25.8 t/s beats M3 Ultra
Agusx1211 · reddit · 2026-08-23
A developer built a custom llama.cpp fork for the M2 Ultra (60 cores, 192GB) that repacks DeepSeek V4 Flash losslessly to 141 GiB — smaller than the public Q4 GGUF (freeing room for context), with byte-identical output and no KV cache quantization.
Key results:
- 25.8 t/s output (42 t/s peak), faster than the 16 t/s seen on M3 Ultra
- SSD KV cache and dynamic lanes, 1M total context across 8 lanes
- Prompt processing is low (250 t/s at 8k-32k) but the SSD cache compensates heavily; the author says compute is still left on the table
Open-sourced at llama-cpp-ds4f-m2-ultra.
More from Infra
- Meta open-sources Muse Spark Open: 816B-parameter model that runs locally on two Mac Studios — gabriberton · 2026-10-16
- Jensen Huang joins BG2 podcast: does AI revenue justify capex, '27 bottlenecks — elonmusk · 2026-10-16
- Mistral's on-device infra rack revealed: DGX Spark, Mac Studio and more — gregmushen · 2026-10-16
- NVIDIA and Sweden's Wallenberg launch Sferical One, a shared AI supercomputer — nvidia · 2026-10-16
- Next infra frontier: self-evolving systems that adapt and self-verify — _shreya_s · 2026-10-16
- Xiaomi's MiMo RL dashboard reveals $3.47M spent on just 30 training steps — bycloud · 2026-10-16