Qwen3.8-27B is now up to ~3× faster on Apple Silicon with mlx-dspark
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
这条记录涉及编程工具或代码能力更新，适合开发者评估工作流变化和可复用价值。
Qwen3.8-27B 在 Apple Silicon 上的运行速度现在提高了约 3 倍，这得益于 mlx-dspark。mlx-dspark 是 DeepSeek 的 DSpark 推测解码草稿器和 z-lab 的 DFlash 的 MLX 移植版本。…
mlx-dspark is an MLX port of DeepSeek's DSpark speculative-decoding drafters (the DeepSpec release), plus z-lab's DFlash, with one lossless verify loop. v0.10.0 adds Qwen3.8-27B via RadixArk's drafter, the first SpecForge/SGLang-packaged head it loads.
Numbers (M4 Pro 48 GB, medians of 3, greedy, output ids identical to plain decoding):
- 8-bit target: 2.45× mean at the auto-picked cap — 3.00× math / 2.38× code / 1.96× chat, 8.3 → 20.3 tok/s (code runs hit 3.18×). Peak ~29 GB.
- 4-bit target: 1.74× at 25.3 tok/s in ~18 GB (same drafter auto-resolves).
- Fun property: 8-bit + drafter (20-27 tok/s) beats plain 4-bit (14.6 tok/s) — 8-bit quality at better-than-4-bit speed.
"Lossless" is checked, not asserted: the target verifies every drafted token, and the Mac app's Race view runs speculative vs plain on the same prompt and diffs the token ids (video is that view).
Everything is pip install mlx-dspark (OpenAI-compatible server + Anthropic Messages API, so it can back Claude Code with a local model), and there's a native Mac app (DMG/Homebrew).