Skip navigation
Reasoning VLAs can think. They just can't think fast. Until now.
0:00 / 1:21

Reasoning VLAs can think. They just can't think fast. Until now.

Zhijian Liu

@zhijianliu10K followers

Visit on X
1,331 likesDownloadOriginal post

192,223 views

Reasoning VLAs can think. They just can't think fast. Until now. Introducing FlashDrive⚡ 🚀 716 ms → 159 ms on RTX PRO 6000 (up to 5.7×) ✅ Zero accuracy loss FlashDrive = streaming inference + DFlash speculative reasoning + ParoQuant W4A8 Real-time reasoning for autonomous driving is here!

32 Comments

Comments from the original post on X (Twitter)

  • Zhijian Liu's profile picture

    @zhijianliu_

    Code + model checkpoints coming soon. The ingredients are already open-source: 🔥 DFlash → ⚡ ParoQuant →

    35 likes

  • Zhijian Liu's profile picture

    @zhijianliu_

    Great work led by @Richard91316073 and @YihaoLiang01, with @hongfeizhang0xF and @jianchen1799.

    13 likes

  • idonteven's profile picture

    @idontev63124802

    Thanks for the paper, I've spent around 2 hours reading it and come up with my response. Screenshot because of twitter character limit.

    13 likes

  • Zhijian Liu's profile picture

    @zhijianliu_

    Thanks for sharing your feedback! The baseline is under BF16.

    5 likes

  • Piotr Mazurek's profile picture

    @tugot17

    awesome work 🫡

    5 likes

  • Crepe Supreme's profile picture

    @crepesupreme

    716ms to 159ms is the line real-time robotics has been waiting on. Every VLA roadmap priced in at least another year before reasoning models could sit inside the control loop. Quietly huge.

    4 likes

  • Josef (atris)'s profile picture

    @atris_eth

    Trying this on an in-car 3090 with a comma soon

    3 likes

  • Neural Drop's profile picture

    @MaoLi92351

    5.7x faster with zero accuracy loss sounds like the kind of magic trick that only works on paper until you actually try it in a car moving at 60mph

    3 likes

  • Henry Zhang's profile picture

    @thehenryinsf

    holy shit. need this deployed somewhere and served on an api right now

    3 likes

  • Don Park's profile picture

    @donpark

    could be more compelling from safety perspective if the model used for speculative reasoning is RL-ed to be more cautious than the full reasoning. safety-first in the literal sense. :-)

    2 likes

  • Brian Chew's profile picture

    @brianchew

    @kstonekuan this looks interesting and in your field!

    2 likes

  • Taro Bushidō's profile picture

    @techietaro

    Sub-160ms VLAs that don't think in molasses. ParoQuant + DFlash cuts latency 5.7x with zero accuracy loss. Lab demos just became production reality. ⚡️

    2 likes

  • Kirito (e/acc) 🏴‍☠️'s profile picture

    @bronzeagepapi

    @yassineyousfi_

    2 likes

  • Avais Aziz's profile picture

    @avaisaziz

    Nice work getting reasoning VLA down to 159ms on RTX PRO 6000 with no accuracy drop. Streaming + DFlash speculative decoding + ParoQuant W4A8 is a clean stack for closing the control loop. Looking forward to the checkpoints.

    2 likes

  • Alen Gabs's profile picture

    @AlenGabs

    very cool !

    1 like

  • Team Reagent's profile picture

    @Reagent_Systems

    Holy smokes this is really cool :O

    1 like

  • Uğur Yekta Başak's profile picture

    @uguryektabasak

    this is very practical

    1 like

  • Atlas AI's profile picture

    @AiAtlas7555

    716ms to 159ms. Real-time VLA reasoning is no longer a lab benchmark.

    1 like

  • Nischay's profile picture

    @NischayJoshi8

    The latency gap is real for live video reasoning. The quality tradeoff at high token budgets is the other side - we found diminishing returns past ~700 thinking tokens for video understanding. Benchmark at

    1 like

  • David Yang's profile picture

    @davidzhaiyang

    Really cool!

    1 like