mlx-dspark
mlx-dspark runs DeepSeek's DSpark and z-lab's DFlash speculative-decoding drafters natively on Apple Silicon via MLX, giving lossless decoding speedups of up to about 4x for supported consumer-size models.
mlx-dspark is a native Apple Silicon MLX implementation of two EAGLE-family speculative-decoding drafters: DeepSeek's DSpark, a semi-autoregressive method originally used to accelerate DeepSeek-V4, and z-lab's DFlash, a block-diffusion method. Both are lossless, since the target model verifies every drafted token, so output matches normal decoding exactly; it does not run DeepSeek-V4 itself, but consumer-size target models such as Gemma-4, Qwen3, Muse-Glimmer, Nemotron, LFM2.5, Ornith-1.0 and Bonsai-27B that have published DSpark or DFlash drafters.
It can be run from the CLI, from Python, or as an OpenAI-compatible API server for tools such as LM Studio, and can drive Claude Code with a model running locally on a Mac. It reports measured speedups per model and per workload (chat, code, math) on an M4 Pro, using a benchmark command with a configurable number of trials.