The successor to January's DFlash predicts a whole block of tokens in parallel and adds a path selector plus a local convolution, squeezing over 20% more accepted output from each verification pass.
2026-08-21 · by Moe Ameen
On August 18, 2026, Inco AI announced DFlash 2, a new version of its speculative-decoding system for running open language models faster. Speculative decoding is a technique that speeds up inference without changing a model's output: a small, fast "draft" model predicts tokens ahead, and the larger "target" model verifies them in a single forward pass, keeping only what it would have produced itself. The result is lossless — the same text as ordinary decoding, generated in fewer, cheaper passes.
DFlash drafts a whole block of tokens in parallel and keeps the top candidates at each position rather than one guess per step. DFlash 2 adds two refinements: 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 addresses "suffix decay," where accuracy drops toward the end of a predicted block. Inco AI reports these yield over 20% more accepted output per verification pass — about 16–25% across benchmarks, a mean acceptance length up roughly 1.05 tokens (21%) over the original DFlash — for around 1% added cycle latency.
In throughput terms, Inco AI cites 2.7–3.4x autoregressive decoding on Qwen3.8-27B and 3.1–4.6x on Meta's Muse Glimmer at batch size 1. DFlash 2 is the successor to DFlash, the parallel drafter Inco AI first released in January 2026 (with roots in 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. At launch, drafters for Qwen3.8-27B and Muse Glimmer are available on Hugging Face, with support across the SGLang, vLLM, llama.cpp, Ollama, and oMLX inference engines. As with any new release, treat the specifics as an early snapshot and confirm them in Inco AI's own materials.
The right way to read a launch like this: the cost of *drafting text* keeps falling, and the cost of *turning text into content people actually see* has not moved at all. DFlash 2 is a genuine engineering advance — running an open model 2.7–3.4x faster with byte-for-byte identical output is real leverage for anyone self-hosting. But it lands entirely on the cheapest step of a content operation. The expensive part — a captioned Short, a brand-exact carousel, a blog, a newsletter, all in one consistent voice, reviewed, and pushed to every platform on a schedule — is exactly what a faster draft does nothing for. That is the part [Kompozy](/) is built to own.
If you run an open model, keep DFlash 2 as your fast front end: draft raw scripts and angles at higher throughput, then drop the best one into Kompozy as a source. Kompozy rewrites it under one [Persona Brief](/glossary/persona-brief) with a banned-word filter and generates 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 fans it across the eight social platforms plus blog and email. Kompozy's own copy runs on managed Claude and OpenAI, so a faster local model is an optional upstream input, not a dependency. DFlash 2 makes drafting cheaper; Kompozy is what turns the draft into finished, on-brand posts and puts them everywhere. For the model-release side of the same trend, see [DiffusionGemma's parallel-decoding launch](/news/diffusiongemma-open-diffusion-llm).
DFlash 2 is a speculative-decoding drafter released by Inco AI on August 18, 2026. It runs an existing open language model faster by having a small draft model predict a block of tokens in parallel, which the larger target model verifies in one forward pass. The output is lossless — identical to ordinary decoding — but produced more cheaply, reaching 2.7–3.4x autoregressive throughput on Qwen3.8-27B.
No. Speculative decoding only changes speed, not quality — the target model verifies every token, so you get exactly the text it would have generated anyway, just faster. DFlash 2 raises how many drafted tokens are accepted per pass (over 20% more than the original DFlash), which lowers cost and latency but does not alter the final output.
Treat it as a background efficiency gain, not a new capability. If you self-host an open model, DFlash 2 makes drafting cheaper; if you use managed tools, faster inference quietly lowers costs over time. Either way, the work that matters — turning a draft into finished, on-brand posts across platforms — is unchanged, and a content engine like Kompozy is what closes that gap.