dflash2

Up to 4× faster LLM decoding on Apple Silicon, lossless. Native MLX port of DeepSeek's DSpark & z-lab's DFlash speculative decoding — Gemma-4, Qwen3.8, Muse-Glimmer, Nemotron, LFM2.5, Ornith-1.0, ternary Bonsai-27B.

Serving Qwen3.8-27B-FP8 on a single DGX Spark (GB10): 7.88 to 58.5 tok/s single-stream from decode strategy alone, weights untouched. Speculative decoding and prefix caching benchmarked, plus DFlash 2 — the only Qwen3.8-27B build that can serve it under vLLM.

Qwen3.8-27B on one DGX Spark GB10: DFlash 2 for 65 tok/s C1 and 120 C4; EAGLE 3/1/4 for ~191 C8, ~330 C16 and ~535 C32; native 262K, pinned SGLang.

Evidence-first lab for speculative decoding, DFlash/DFlash2, block diffusion, exactness, acceptance, and inference performance

Qwen 3.8 DFlash2: 27B AI Running at 236 TOK/S Locally! - High-speed local prose and code autocomplete powered by speculative decoding with lightweight draft models.

Qwen3.8-27B + DFlash2 block-diffusion speculative decoding on Apple Silicon: 809 tok/s, prefix-cached agentic serving at 262k context

Native macOS control center and local AI agent gateway for Qwen3.8 on Apple Silicon — MLX, DFlash2, OpenAI Responses, Anthropic Messages, Claude Code, Codex, OpenCode and Grok Build.

DFlash2 extension of MiaAI-Lab's one-DGX-Spark Qwen3.8 NVFP4 SGLang recipe

Blackwell-optimized llama.cpp bundle with DFlash2, MXFP6/MXFP8/NVFP4, TurboQuant KV, and GPU-resident speculative handoff.

Qwen3.8-27B on SGLang for Blackwell SM120: DFlash2 block-16 + triton. 268 tok/s on one RTX PRO 5000 (48GB), full 262K context. Measured, with honest sample sizes.

Standalone eugr/spark-vllm-docker recipes for Qwen3.8-27B NVFP4 and official FP8 on one DGX Spark with dflash2