Inco AI's speculative-decoding drafter that predicts a whole block of tokens at once so a target model can verify them in a single pass — 2.7–3.4x the throughput of autoregressive decoding on Qwen3.8-27B, with identical output. Announced August 18, 2026.
Last verified · 2026-08-21 · by Moe Ameen
DFlash 2 is a speculative-decoding drafter from Inco AI, announced on August 18, 2026. It is not a chatbot or a model you talk to on its own — it is the piece of plumbing that makes an existing open language model generate text faster without changing what it says. In speculative decoding a small, fast "draft" model guesses ahead, and the larger "target" model verifies those guesses in a single forward pass; whatever the target model accepts is output. Because the target model has the final say, the result is lossless — you get the exact same text you would from ordinary decoding, just produced in fewer, cheaper passes.
The DFlash approach drafts a whole block of tokens in parallel and keeps the top candidates at each position, rather than committing to one guess per step. DFlash 2 sharpens that with two changes: a lightweight "path selector" that scores adjacent token pairs to trace coherent sequences through the candidate lists instead of taking each position's top pick independently, and a "local convolution" (two-tap dynamic depthwise convolutions) that fixes "suffix decay," the tendency for accuracy to fall off toward the end of a predicted block. Inco AI reports this yields over 20% more accepted output per verification pass — roughly 16–25% across benchmarks, a mean acceptance length up about 1.05 tokens (21%) over the original DFlash — for around 1% added cycle latency.
In practice that lands as 2.7–3.4x the throughput of autoregressive decoding on Qwen3.8-27B and 3.1–4.6x on Meta's Muse Glimmer (batch size 1). It is the successor to DFlash, the parallel drafter Inco AI first released in January 2026 (originating from work at Z Lab), which reached close to 3x autoregressive speed and whose models have been downloaded over 3.5 million times on Hugging Face. Drafters for Qwen3.8-27B and Muse Glimmer shipped alongside the launch, with support across the SGLang, vLLM, llama.cpp, Ollama, and oMLX inference engines. Treat the exact numbers as an early snapshot and confirm them in Inco AI's own materials before you depend on one.
The important thing to understand about DFlash 2 is that it changes *how fast* your model writes, not *what* you can make. Speculative decoding is lossless by design — the text is byte-for-byte what the target model would have produced anyway, just cheaper. So it moves one number: the cost and latency of the drafting step. That step was never the hard part of a content operation. Turning a draft into a captioned Short, a brand-exact carousel, a blog, and a newsletter, keeping the voice consistent, reviewing it, and getting it onto every platform — that is the 95% DFlash 2 doesn't touch. [Kompozy](/) owns exactly that 95%.
So the clean way to use them is stacked, not either-or. If you self-host, run DFlash 2 to speed up your open model and spin out raw scripts and angles at higher throughput, then drop the best one into Kompozy as a source. Kompozy rewrites it under a single [Persona Brief](/glossary/persona-brief) with a banned-word filter and fans it into a full batch — captioned [Persona Shorts](/glossary/persona-shorts) and avatar video, brand-exact [Carousels](/glossary/hyperframes), Photo Posts and Quote Graphics, a Blog Article, and an Email Newsletter — each passing a per-post review gate before [Autopilot](/glossary/autopilot) schedules and publishes across the eight social platforms plus blog and email. If you would rather skip the GPU entirely, Kompozy already runs managed Claude and OpenAI for its copy, with a bring-your-own-key option on the Founding tier. DFlash 2 makes the raw words cheaper; Kompozy is what turns them into finished, on-brand content and puts it everywhere. The same pattern works with any fast open model — see [DiffusionGemma](/ai-tools/diffusiongemma) or a [local Gemma 4 26B engine](/ai-tools/gemma-4-26b-local-engine).
DFlash 2 is a speculative-decoding drafter from Inco AI, announced August 18, 2026. It speeds up an existing open language model by having a small draft model predict a whole block of tokens at once, which the larger target model verifies in a single forward pass. The output is identical to ordinary decoding — it is lossless — but produced in fewer, cheaper passes, reaching 2.7–3.4x the throughput of autoregressive decoding on Qwen3.8-27B.
No. Speculative decoding is lossless: the target model verifies every drafted token, so you get exactly the text that model would have generated anyway, just faster. DFlash 2's improvements — a path selector and a local convolution that fixes suffix decay — increase how many drafted tokens get accepted per pass (over 20% more), which raises speed without altering the final output.
At launch Inco AI released DFlash 2 drafters for Qwen3.8-27B and Meta's Muse Glimmer on Hugging Face, with support across the SGLang, vLLM, llama.cpp, Ollama, and oMLX inference engines. It is a component you add to a serving stack, not a standalone app, so you also need the target model it drafts for.
DFlash 2 makes the drafting step of an open model faster and cheaper, but drafting is only the start of making content. Kompozy takes that raw text and generates finished, on-brand assets — captioned Persona Shorts, carousels, quote graphics, a blog, and a newsletter under one Persona Brief — then reviews and publishes them across the eight social platforms plus blog and email. Use DFlash 2 to draft fast; use Kompozy to finish and distribute.