NVIDIA Unveils VFlash: A System-1 Parallel Decoding Framework for Physical AI

NVIDIA researchers (the LPR Team and Jim Fan's GEAR Lab, among others) have released VFlash, a general decoding framework for Physical AI positioned as a "System-1 visual model": instead of forcing embodied models to think only in text, it lets them think with "visual planning" — parsing the visual world in parallel at high speed to directly serve action execution.

Confirmed

Core idea: outputs of most embodied prediction tasks are fixed-length float vectors (e.g., bounding boxes are always 6 tokens, a 10-waypoint robot arm trajectory is 54 tokens), so VFlash decodes them all at once in parallel rather than token by token. Jim Fan reported 5.1x faster robot inference with 61% lower error; follow-up posts from the team cite up to 5.7x speedups with full-stack support.
Decoding coverage: outputs span both text and structured geometric outputs; Zhiding Yu describes it as a universal structured parallel decoder for Physical AI.
Engineering: VFlash gets inference acceleration across NVIDIA's full stack — Jetson Thor, DGX Spark, and H100 — built on the TensorRT-Edge-LLM and TensorRT stacks, targeting robotic manipulation and autonomous driving.
Downstream benefits: VLM features can be frozen so a world action model (WAM) or diffusion action head links to VFlash's visual primitive tokens via cross-attention; fast, accurate visual planning predictions bring multiple downstream gains.

Why it matters

VFlash complements "System-2 text reasoning" with a "System-1 visual planning" route, compressing visual prediction directly into fixed-length vector outputs and bypassing the autoregressive decoding bottleneck.
End-to-end deployment on edge devices like Jetson Thor means the framework isn't just a paper concept — it can run real-time inference for robotic manipulation and autonomous driving.

2026-11-06 ~ 2026-11-06 · 10 related posts

Primary sources

VFlash decodes robot actions in one shot: 5.1x faster inference, 61% lower 3D trajectory error — DrJimFan · 2026-11-06 00:31 (UTC+08:00)
NVIDIA unveils VFlash: parallel decoding of visual plans for physical AI — ZhidingYu · 2026-11-06 00:13 (UTC+08:00)
VFlash brings 5.7x faster visual planning inference on Jetson Thor, DGX Spark and H100 — ZhidingYu · 2026-11-06 01:45 (UTC+08:00)

[source] NVIDIA unveils VFlash: parallel decoding of visual plans for physical AI — ZhidingYu · 2026-11-06
[source] VFlash decodes robot actions in one shot: 5.1x faster inference, 61% lower 3D trajectory error — DrJimFan · 2026-11-06
NVIDIA's Jim Fan unveils VFlash: a System 1 vision model for Physical AI — andrew_n_carr · 2026-11-06
NVIDIA's VFlash: a generalist parallel decoder for structured Physical AI outputs — ZhidingYu · 2026-11-06
VFlash finding: lightweight drafter can beat its autoregressive target on decoding accuracy — ZhidingYu · 2026-11-06
VFlash design keys: primitive boundary alignment and position alignment for parallel decoding — ZhidingYu · 2026-11-06
Frozen VLM features plus VFlash planning boost world action models on MolmoSpace — ZhidingYu · 2026-11-06
[source] VFlash brings 5.7x faster visual planning inference on Jetson Thor, DGX Spark and H100 — ZhidingYu · 2026-11-06
NVIDIA's VFlash: fast visual planning VLM for physical AI, 5x faster on Jetson Thor — ZhidingYu · 2026-11-06

1 near-duplicate retellings: ZhidingYu