DeepSeek DSpark Speculative Decoding Ported to Mac with Lossless Speedup
A-Rahim · reddit · 2026-07-07
Developers have ported DeepSeek's speculative decoding drafter, DSpark (built for DeepSeek-V4), to MLX, enabling native Mac execution alongside an OpenAI-compatible service and benchmarks. As an EAGLE-style drafter, the main model still verifies every drafted token, ensuring the output perfectly matches standard decoding (byte-for-byte identical in greedy mode, and verified exact sampling in temperature mode). It currently supports Qwen3 4B/8B/14B and Gemma-4 12B.
Benchmarked on an M4 Pro with 8-bit instruction targets, it achieves a 1.4–1.6x speedup for single users, and up to roughly 2x for code/math tasks using Gemma. The author notes that Mac speeds won't hit the commonly cited 2–4x figures, which the original paper never actually claimed.
More from Infra
- Encrypted GPT-2 speedrun challenges FHE inference: best is 3.031 s/token, goal sub-1s — tensorqt · 2026-10-27
- Open-source Latch compiles PyTorch into training megakernels, 2.75x faster on H100 — sguada · 2026-10-27
- NVIDIA claims HGX Rubin NVL8 delivers 10x tokens per megawatt vs B200 — thursdai_pod · 2026-10-27
- NVIDIA Claims HGX Rubin NVL8 Delivers 10x More Tokens per Megawatt Than B200 — thursdai_pod · 2026-10-27
- Together lets multiple LoRA post-training experiments share the same GPU capacity — togethercompute · 2026-10-27
- Commoditization dilemma: either human labor or GPUs become the commodity — _AustinCalvert_ · 2026-10-27