DFlash 2 brings 20% faster parallel drafting to the inference stack

Share
DFlash 2 brings 20% faster parallel drafting to the inference stack

Inference is eating AI's bill, and a new open-source release just made it cheaper to run.


DFlash 2, a block-diffusion draft model from Inco AI, ships checkpoints for Qwen3.8-27B and Muse-Glimmer-30B — and claims over 20% more output per verification pass at roughly 1% added latency. The core idea is speculative decoding: a small draft model guesses a block of tokens, the large model verifies the whole block in one forward pass. DFlash made that guess itself parallel (one pass, not one token at a time); DFlash 2 pushes it further by keeping each position's top candidates and tracing one coherent path through them. Across benchmarks the acceptance-length gain runs 16–25%, and on Qwen3.8-27B it hits the throughput of autoregressive decoding at batch size one in SGLang — meaning you get more tokens per second without burning extra GPU cycles. The release lands at a moment when agents are the dominant consumer of inference tokens: they read, plan, and call tools for hours, and every token costs a full forward pass. NVIDIA measured DFlash 2 on Blackwell GPUs; Google reported three times more tokens per second on TPUs with DFlash; CoreWeave's production Kimi K2.7 Code endpoint runs DFlash by default. On Hugging Face, DFlash models have passed 3.5 million total downloads. The checkpoints are live now — DFlash 2 drafts for Qwen3.8-27B and Muse-Glimmer-30B on Hugging Face, with integration in SGLang, vLLM, llama.cpp, and oMLX. For teams running local or self-hosted models, the math is simple: more tokens per verification pass means lower cost per agent-hour, which is the number that matters as workloads shift from chat to long-running tasks.

What to watch: whether DFlash 2's gains hold up as models scale past 30B parameters, and how quickly inference engines bake it into their default paths.

Does speculative decoding change how you think about running agents locally? Write to us via our contact page.

Sources: DFlash 2 GitHub (z-lab)

Get SiliconNoon in your inbox

One email a day with the AI stories that matter — what happened, the numbers, and why it matters. No spam, unsubscribe in one click.

Read more

DeepSeek open-sources Harness — everything is a plugin

DeepSeek open-sources Harness — everything is a plugin

DeepSeek is moving up the stack from models to the software developers actually use, and Uber is scaling robotaxis across Europe. Two open-source and mobility moves to start your Friday. DeepSeek released DeepSeek Harness (dsh), an open-source agent harness under the MIT license in developer preview, built on an architecture where "every capability is a plugin" — from the model adapter and tool registry to the session log and the agent loop itself.

Minitap says Google's Artemis shipped its code with the credit stripped

Minitap says Google's Artemis shipped its code with the credit stripped

Google's Pixel test team open-sourced an Android automation agent this week. The team it was built on top of says the authorship trail was rewritten before anyone saw it. Minitap says Google's Artemis contains its code, and that the names of the people who wrote it were removed. Minitap published mobile-use in August 2025 under Apache 2.0 — an agent that drives real Android and iOS apps, 2,821 GitHub stars, and the reference implementation behind a clutch of mobile-agent papers.

Hugging Face runs async RL training with no NCCL and no weight ship

Hugging Face runs async RL training with no NCCL and no weight ship

The hardest part of reinforcement learning is not the algorithm. It is getting a trainer and a bunch of inference servers to keep passing weights back and forth without one of them stalling. Hugging Face just published a recipe that does it with three rented jobs, one storage bucket, an HTTP proxy, and no NCCL at all. Hugging Face trained a reasoning model end to end using async GRPO and LoRA, and the whole weight sync is a file upload.

Fastino's 340M decision model beats its own 1B on routing

Fastino's 340M decision model beats its own 1B on routing

Most agent stacks answer a yes/no question by having a large model write a paragraph. Fastino's answer is to not ask a large model at all — and its new release is small enough to run on the CPU you already have. Fastino shipped GLiNER2.5-Decide, a 340M-parameter classifier that takes a bounded question — intent, routing, sentiment, severity, spam — and returns a label in a single forward pass, with no prompt template and no generated tokens.

Block open-sources Berd, a desktop app for juggling AI agents across models

Block open-sources Berd, a desktop app for juggling AI agents across models

Block just dropped Berd into the open — a desktop app that lets you work with AI agents from any provider in a single window, released under Apache 2.0. The project, built on Tauri 2 and React 19, wraps Block's own Goose agent backend and already has 216 GitHub stars within days of going live. The bet is straightforward: as teams start using multiple AI agents for different tasks, they need one interface instead of five browser tabs.