NVIDIA's VFlash decodes robot actions 12x faster
The model skips token-by-token decoding because it already knows how long the answer will be.
NVIDIA researchers have released VFlash, a vision model that spits out a full robot trajectory or a bounding box in a single step instead of one token at a time. Jim Fan of NVIDIA's GEAR Lab says it delivers up to 12x faster action decoding and a 5.1x overall robot inference speedup on edge GPUs like the Jetson Thor and DGX Spark.
every coordinate is written in ink and every later point loses the chance to backspace
The trick is simple once you see it. Most embodied prediction tasks produce fixed-length float vectors, so the model already knows how long the answer is before it starts writing. A bounding box is always 6 tokens. A 10-waypoint robot arm trajectory is always 54. If the length is known, there is no reason to generate it autoregressively (one token at a time, each conditioned on the last).
Faster and more accurate, which is the unusual part
Speedups normally cost you accuracy. Fan says this one does not. VFlash reportedly ends up more accurate than the VLM backbone it sits on, with up to 61% lower 3D trajectory error. His explanation is that one-shot output dodges the cascading errors of token-by-token decoding, where an early mistake poisons everything that follows.
That framing is the most quotable thing in the announcement, and it is also the clearest argument for why parallel decoding might be more than a latency hack. Fan calls it a System-1 vision model, the reflex layer rather than the deliberate one.
Shown versus claimed
Every number here comes from the team that built the model. There is no outside benchmark yet, and the 12x and 61% figures are both stated as upper bounds, not averages. Zhiding Yu, the project lead, describes it as joint work between NVIDIA's LPR Team and GEAR Lab with collaborators including @YuCheng_Chou_, Shihao Wang and Jin Huang.
The mitigating detail is that the paper and the code are both out. Fan says it is open paper, open source, which means anyone with a Jetson Thor can check the inference numbers themselves rather than taking the thread's word for it. Given how often robotics speed claims arrive as a video and nothing else, that is the part worth paying attention to.
Embodied models don't just think in text, they can think in visual plans, and fast.
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.