DeepSeek DSpark Technology Ported to Apple Silicon, Boosts Local Mac LLM Speed by Up to 60%

DeepSeek's speculative decoding technology DSpark, open-sourced on June 27, has been ported to Apple Silicon (Mac) by engineer Abdur Rahim under the project name mlx-dspark. The port supports Gemma-4 12B and Qwen3-4B models, achieving speedups of approximately 1.6× (from 18.4 tok/s to ~30 tok/s) and 1.4× (from 52.9 tok/s to ~73 tok/s) on M4 Pro, respectively.

Technical Principles and Implementation

• DSpark core idea: A small draft model quickly generates candidate tokens, which are then batch-verified by the target model. If accepted, they are kept; otherwise, regeneration occurs.
• Cost differences on Apple Silicon: Data center GPUs have fixed batch verification costs, while Apple Silicon's verification cost scales linearly with the number of candidate tokens. Rahim measured that each additional token verification for Gemma-4 12B takes about 14 ms, and built a cost model yielding a theoretical speedup upper bound of ~2.2×.
• Implementation details: The draft model is extracted from HuggingFace checkpoints and quantized to 4-bit (only 1.8GB). The target model defaults to 8-bit precision (bf16 verification is more expensive and actually slower).

High-Fidelity Reproduction and Sampling Support

• Most local ports only support greedy decoding, but mlx-dspark implements the temperature sampling method from the paper, producing output distributions strictly identical to the target model, byte-for-byte.
• Rahim found: When the draft model is paired with a non-instruction-tuned base target model, the candidate acceptance rate is only 47%; switching to an instruction-tuned version raises the acceptance rate to 82%.

Integration of DFlash Scheme

• At the request of DFlash paper author Jian Chen, Rahim integrated the DFlash scheme into mlx-dspark. DFlash uses parallel block diffusion (generating 16 tokens at once), achieving a speedup of ~2.1× (~36 tok/s) on code and math tasks, outperforming DSpark.
• However, in open-ended chat scenarios, DFlash's acceptance length is limited, making DSpark faster. mlx-dspark v0.0.3 allows users to manually adjust DFlash's effective block length to suit different tasks.

Impact and Outlook

• This is the first native Apple Silicon implementation of DSpark since its open-source release, enabling Mac users to enjoy acceleration without relying on data center GPUs.
• Rahim indicates the method can scale to larger draft models (e.g., Qwen3-8B and 14B).
• Concurrently, DeepSeek is actively recruiting, including Tsinghua University PhD student Gu Yuxian (Apple PhD Scholar), whose research covers model compression and efficient architectures, synergizing with technologies like DSpark.