TensorFold: exact speculative decoding on Apple Silicon. Drafted output is byte-identical to serial decoding, yet measured substantially faster on real Mac hardware.
• M5 Max (128 GB), Nemotron 3.5 Lightning 4-bit: 188–206 tok/s vs ~138 tok/s for mlx_lm (short answers with thinking). Holds 175–162 tok/s at 20k–60k context.
• Qwen3.8-27B 4-bit + DFlash2 on M5 Max: ~189 tok/s on code vs ~26 tok/s without drafts. Independent tests show 2.5–3.7× over other engines using the same drafter.
• M5 Ultra (256 GB) community runs (Qwen3.8 Flash Next variants): 148–189+ tok/s decode across 1k–64k+ contexts; prefill often 2k–3k+ tok/s. Concurrent 8-stream Qwen3.8-27B: ~1.84× vs llama.cpp (~160 vs ~87 tok/s).
• Two M5 Ultra Mac Studios over Thunderbolt 5 (speed-up/split): ~1.2–1.5× decode and 1.7–1.8× faster prefill from ~8k tokens on large models (e.g. Flash Next), remaining token-identical. Prefill can reach several thousand tok/s combined.
• Nemotron variants have hit 200+ tok/s peaks in-engine with MTP; prefill patches have delivered ~10% gains on M5 Ultra while preserving exactness.These numbers turn Mac Studio’s unified memory advantage into usable, deterministic local inference speed.
There is too much to list, but a formal conversation really needs to start to take this where it can go.
https://t.co/aHePPbMYUk
@shshnkp@ashxhart
Microsoft released a 4B open-weight AI that uses your computer. It scores 66% on OSWorld-Verified. Humans: 72.4%
In Microsoft's demo, one spreadsheet task took 39 seconds. A 27B model, clicking only, needed 9 min 56 s
Send this to the friend who lives in Excel.