Open-Source mlx-dspark Boosts LLM Inference on Mac by 3.3x

A-Rahim · reddit · 2026-08-13

Developer A-Rahim released the open-source project mlx-dspark, which leverages speculative decoding on Apple Silicon to significantly accelerate Meta's Muse Glimmer 30B model.

In real-world testing on an M4 Pro chip, the 8-bit model's generation speed increased from 8.2 tok/s to 18-26 tok/s. Math tasks saw the best improvement (3.27x), followed by code (2.5x) and chat (2.22x). Because the target model verifies every token, the output is byte-identical to normal decoding, meaning there is no quality trade-off.

Additionally, the 4-bit quantized run achieves 25 tok/s (1.7x speedup) requiring only 18GB of memory, while the 8-bit run peaks at 40GB. This effectively delivers 8-bit quality at 4-bit speeds.

Original post →

More from Infra

Infra channel →