Humanizer's DSpark draft reports 1.54x rewriting throughput on Apple silicon

A faster draft for a specific rewriting model

Humanizer has a new downloadable DSpark draft aimed at accelerating local text rewriting on Apple silicon. The author reports 33.49 tokens per second against 21.71 for the target model alone in a short-output MLX pilot, roughly 1.54 times the throughput.

This is a target-specific speculative draft: it proposes tokens, and the Humanizer target verifies them. It requires that target model and is not a standalone Humanizer fine-tune. The October 15 update follows the initial October 14 weights release and additional training documented by the author.

The comparisons matter. The original DFlash draft reached 23.43 tokens per second in the same reported run. The previous DSpark checkpoint reached 32.63, so the new checkpoint’s gain over its own predecessor was about 2.6%, not 54%.

What the pilot actually measured

The author used an Apple M5 Pro, a mixed 4/8-bit target with FP32 floating computation, and 4-bit drafts. The test used ten held-out inputs with two repeats per model, greedy decoding, a draft cap of three and at most 64 output tokens. Median throughput includes prefill and verification. Across eight arms, including previous checkpoints, all 160 requests reportedly matched target-only tokens.

Those are author-reported measurements, not an AITrending reproduction. The evaluation explicitly notes substantial trial variability and does not claim a robust or long-form speedup. Matching tokens in this pilot also does not establish that every rewrite preserves every factual detail.

The download comes with a research runtime

The documented route uses Python 3.12 and pinned MLX packages with a patch to the specified mlx-dspark revision. Users download run_mlx.py from the checkpoint repository and pass text to it; --check also runs the target alone and fails on a token mismatch. The runner downloads the pinned target, roughly 9 GB, plus draft weights.

The existing MLX runtime was tested, but the fresh-install recipe was not rerun in a clean environment. vLLM and SGLang remain unverified for this checkpoint, with no CUDA serving or speed claim. The checkpoint is Apache-2.0 licensed and retains the upstream DFlash MIT notice.