# Inco AI releases DFlash 2, reports 21% longer accepted token sequences > Source: > Published: 2026-08-19 21:25:25+00:00 # Inco AI releases DFlash 2, reports 21% longer accepted token sequences **Inco AI adds a path selector and local convolution to DFlash 2, reporting 21% longer accepted drafts with 1.3% added cycle latency.** By [RuntimeWire Staff](/author/runtimewire-staff) ยท Published Primary source: [Inco AI](https://inco.ai/blog/dflash2/) ## Why it matters Agent workloads multiply token demand and serving costs. If Inco AI's gains hold across production stacks, DFlash 2 can reduce target-model passes without changing model output. Zhijian Liu is a UC San Diego-affiliated academic author behind DFlash, the parallel speculative-decoding method that [Inco AI](https://inco.ai/?ref=runtimewire) released in an updated version on August 18. DFlash 2 adds two small components designed to get more accepted tokens from each expensive LLM verification pass. [Inco AI reports](https://inco.ai/blog/dflash2/?ref=runtimewire) a 21% increase in average acceptance length over the original DFlash. Its published Qwen benchmark uses a five-layer Qwen3-4B DFlash model on GSM8K. The path selector and convolution responsible for the reported gain added 1.3% to draft-verify cycle latency in Inco AI's tests. Inco AI says the final output remains unchanged because the larger target model still verifies every proposed token. The DFlash paper names [Liu](https://datascience.ucsd.edu/people/zhijian-liu/?ref=runtimewire) alongside UC San Diego-affiliated authors Jian Chen and Yesheng Liang. The paper was submitted to [arXiv](https://arxiv.org/abs/2602.06036?ref=runtimewire) in February. Their central move was to make the drafting stage parallel, replacing a small model's token-by-token guesses with an entire proposed block generated in one pass. Inco AI is building its first public product around the DFlash approach. Inco AI describes DFlash 2 as the first piece of an end-to-end inference stack built for agent workloads, where a single task may run for hours and generate far more tokens than a conventional chat session. ### A parallel draft learns to read its own guesses Speculative decoding uses a smaller model to propose several future tokens before asking the larger target model to check them together. Accepted guesses save forward passes. Rejected guesses are discarded, preserving the target model's output. Most speculative-decoding systems still generate the draft sequentially. DFlash proposes every position in parallel, which removes that drafting loop but creates a different problem: independently plausible tokens can make a poor sequence when placed next to one another. DFlash 2 adds a path selector that keeps the top 16 candidates at each position, scores neighboring token pairs and walks through those precomputed scores to choose a coherent sequence. [Inco AI says](https://inco.ai/blog/dflash2/?ref=runtimewire) the selector adds 2 million parameters and 0.6% cycle latency. In Inco AI's Qwen3-4B test, it produced longer accepted drafts than a DSpark correction module while using roughly 40 times fewer parameters. The second addition targets what Inco AI calls suffix decay. The original drafter becomes less accurate near the end of a proposed block, reducing the number of tokens that survive verification. DFlash 2 inserts a local two-tap convolution around each attention and feed-forward sublayer, allowing every position to mix information with its immediate predecessor while retaining parallel computation. [Inco AI says](https://inco.ai/blog/dflash2/?ref=runtimewire) the convolution adds 16.5 million parameters, about 3% of the five-layer drafter, and 0.7% cycle latency. Inco AI's experiments found that this small local operation recovered much of the end-of-block accuracy obtained by tripling the drafter from five layers to 15, an expansion that carried a 15.2% latency cost. [Across Inco AI's benchmark suite](https://inco.ai/blog/dflash2/?ref=runtimewire), the selector and convolution improved average acceptance length by 21% over DFlash, with individual gains ranging from 16% to 25%. The added draft-verify cycle latency was 1.3%. Those are Inco AI's benchmarks, and the headline throughput figures depend on model, task, sampling settings, hardware and concurrency. For the newly released [Qwen3.8-27B drafter](https://huggingface.co/incoai/Qwen3.8-27B-DFlash2?ref=runtimewire), [Inco AI reports](https://inco.ai/blog/dflash2/?ref=runtimewire) 2.7x to 3.4x the throughput of autoregressive decoding in SGLang at batch size one. Its [Muse Glimmer drafter](https://huggingface.co/incoai/Muse-Glimmer-30B-DFlash2?ref=runtimewire) produced 3.1x to 4.6x throughput in Inco AI's tests. ### The integrations are doing the selling The original [open-source DFlash repository](https://github.com/z-lab/dflash?ref=runtimewire) is supported in SGLang, vLLM, TensorRT-LLM and llama.cpp. The DFlash 2 release includes configurations for SGLang, vLLM, llama.cpp and an [oMLX build for Apple Silicon](https://github.com/z-lab/omlx-fork/releases/download/0.6.2-dflash2/oMLX-0.6.2-zlab-dflash2-arm64-signed.dmg?ref=runtimewire). Independent hardware vendors have also tested the first DFlash. [NVIDIA reported](https://developer.nvidia.com/blog/boost-inference-performance-up-to-15x-on-nvidia-blackwell-using-dflash-speculative-decoding/?ref=runtimewire) up to 15x higher throughput for [gpt-oss-120b](/models/openai/gpt-oss-120b) on an eight-GPU Blackwell system at the same interactivity target. [Google reported](https://developers.googleblog.com/en/supercharging-llm-inference-on-google-tpus-achieving-3x-speedups-with-diffusion-style-speculative-decoding/?ref=runtimewire) a 3.13x average speedup for a standalone JAX DFlash benchmark on TPU v5p. Its vLLM TPU-pipeline comparison showed a 2.29x end-to-end speedup over autoregressive decoding. Those results validate the original architecture on different accelerators. They do not independently confirm DFlash 2's new 21% acceptance-length claim. [CoreWeave](https://www.coreweave.com/blog/kimi-k2-7-code-now-available-on-serverless-inference-with-leading-benchmark-price-performance?ref=runtimewire) uses DFlash by default for its [Kimi K2.7 Code](/models/azure/kimi-k2.7-code) endpoint, while [Inco AI says](https://inco.ai/blog/dflash2/?ref=runtimewire) NVIDIA, Red Hat, Modal, Meta, Poolside and Xiaomi have published or shipped compatible drafters. Inco AI also reports that DFlash models passed 3.5 million downloads on Hugging Face by August 2026. DFlash is distributed through inference engines and model repositories that developers already use, giving Inco AI a path into existing serving stacks while it builds a broader inference product around the method. The timing follows the workload. Agents that plan, call tools and revise outputs turn inference efficiency into an operating expense rather than a responsiveness feature. Every additional accepted token reduces the number of full target-model passes needed to finish the task. DFlash 2 tries to lower that cost by improving the parallel drafter while retaining the target model's verification step.