{"version":1,"type":"story","url":"https://digestai.news/story/inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17","json":"https://digestai.news/story/inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17.json","markdown":"https://digestai.news/story/inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17.md","slug":"inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17","headline":"Inco AI’s 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).","keyPoints":["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"],"whyItMatters":"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.","category":{"slug":"research","name":"Research","url":"https://digestai.news/category/research"},"entities":{"companies":["Inco AI","Qwen"],"models":["Qwen3.8-Flash-Next","DFlash2","Flash-Next"],"people":[]},"firstPublishedAt":"2026-10-30T17:19:41Z","updatedAt":"2026-10-30T17:19:41Z","sourceCount":1,"hasPrimarySource":true,"sources":[{"outlet":"huggingface.co","title":"PSA: Upgrade Flash Next to DFlash2 if you haven't! +17% on average","url":"https://huggingface.co/nazeshinjite/Qwen3.8-Flash-Next-DFlash2-GGUF","publishedAt":"2026-10-30T17:19:41Z","type":"primary","primary":true,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1xjlgm4/psa_upgrade_flash_next_to_dflash2_if_you_havent/","points":null}],"thread":null,"cite":{"text":"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","publisher":"Digest AI","title":"Inco AI’s DFlash2 drafter boosts Qwen3.8-Flash-Next speed by 17%","datePublished":"2026-10-30T17:19:41Z","url":"https://digestai.news/story/inco-ais-dflash2-drafter-boosts-qwen3-8-flash-next-speed-by-17"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"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"}