# Inco AI’s DFlash2 drafter boosts Qwen3.8-Flash-Next speed by 17%

Digest AI · Research · published 2026-10-30T17:19:41Z

Canonical: https://digestai.news/story/inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17

## Summary

Inco AI released DFlash2, an updated speculative decoding drafter for Qwen3.8-Flash-Next, offering a **17% average speedup** over the original Flash-Next model. The drafter proposes token blocks in parallel and uses a sliding-window attention mechanism (2048 tokens) to reduce compute overhead. Benchmarks on an Apple M5 Max show DFlash2 achieves **75.8 tokens/second** at its default settings, compared to **74.5 t/s** for Flash-Next’s baseline, with minimal trade-offs in output quality. The GGUF conversion of DFlash2 is **6x smaller** than Flash-Next’s MTP head on disk, at **287 MiB** (Q4\(0\) quantization), while runtime costs remain comparable due to a larger compute buffer (2.9 GiB).

## Key points

- DFlash2 drafter improves Qwen3.8-Flash-Next speed by **17%** on average, measured at **75.8 tokens/sec** vs **74.5 t/s** for Flash-Next
- GGUF file size drops to **287 MiB** (Q4\(0\)) from **6x larger** MTP head, with no runtime cost increase
- Optimized for list-heavy workloads: deeper drafts (+48% speedup) hurt prose quality slightly but help structured outputs

## Why it matters

For developers running Qwen3.8-Flash-Next on edge devices, DFlash2 cuts inference latency without sacrificing accuracy, enabling faster local AI processing. The sliding-window attention and block-diffusion approach could reduce cloud costs for high-throughput tasks like code generation or list-based reasoning.

## Sources

1. [PSA: Upgrade Flash Next to DFlash2 if you haven't! +17% on average](https://huggingface.co/nazeshinjite/Qwen3.8-Flash-Next-DFlash2-GGUF) (huggingface.co, 2026-10-30, primary source)

## Cite

Digest AI, "Inco AI’s DFlash2 drafter boosts Qwen3.8-Flash-Next speed by 17%", 30 October 2026, https://digestai.news/story/inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17

---

Written by Digest AI's editorial model from the linked sources; the sources are the record. Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse
JSON: https://digestai.news/story/inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17.json