Nvidia’s VFlash decodes robot trajectories in one pass, 5.1x faster

Nvidia’s Jim Fan announced VFlash on X, a vision model that outputs whole geometric predictions in parallel instead of token by token. He says it decodes actions as much as 12x faster and speeds up whole-robot inference 5.1x on small, on-device Nvidia GPUs.

One pass for outputs whose length is known

Fan’s argument is that most embodied prediction tasks produce fixed-length vectors of numbers. Object localization, robot trajectory planning and dense 3D motion all fall in that group. In his examples, a bounding box always takes 6 tokens and a robot arm path of 10 waypoints takes 54, so both sizes are known before decoding starts. A standard vision-language model still emits them one token at a time. VFlash instead outputs the entire shape at once.

According to Fan, the speedups were measured on Jetson Thor and DGX Spark, two Nvidia platforms for running models locally. He gives no model sizes, task benchmarks or latency in milliseconds in the post.

He also says VFlash is more accurate than the VLM backbone it builds on, with up to 61% lower 3D trajectory error. His explanation is that autoregressive decoding compounds mistakes: each coordinate is written first and cannot be revised, so every later point inherits the error. Decoding the whole shape together avoids that cascade.

Fan sums up the pitch this way: “Everything that moves will be autonomous, and everything autonomous will need to think on its own chip.” He adds that the work is an open paper and open source, with links in a follow-up post in the thread. We could not open the thread, so the paper, code and license are not reviewed here.

Why the edge hardware makes this trick matter

Token-by-token decoding is especially costly on edge devices. Jetson AGX Thor, according to a RidgeRun technical overview, pairs a Blackwell GPU rated at 2,070 FP4 TFLOPS with 128 GB of LPDDR5x memory at 273 GB/s of bandwidth. Generating each token means streaming model weights through that memory, so a modest bandwidth figure caps how fast a long sequence of tokens can be produced. Cutting the number of sequential steps attacks that limit directly.

It is also a different route from speculative decoding, which Nvidia already supports. Its TensorRT Edge-LLM benchmarks for Jetson AGX Thor list EAGLE3 on Qwen3-8B at 153.7 tokens per second against 47.3 for vanilla decoding at batch size 1, about 3.25x by our arithmetic from the table. At batch size 8 the same setup falls to roughly 1.58x. Speculative methods still generate text sequentially and gain less under load, while VFlash targets outputs whose structure is fixed.

What would confirm the 5.1x claim

The figures are Nvidia-side claims from a social post and have not been independently verified. Several details will decide how far they travel:

• The baseline. A 5.1x overall gain depends on what “overall” includes. Vision encoding and language steps still take time, which is why the action-decoding figure is larger than the end-to-end one.
• The accuracy comparison. “Up to 61%” is a best case. The paper’s tables should show the range across tasks and datasets.
• Scope. The method fits outputs of known length. Open-ended reasoning or planning text, which robot models also produce, would still need ordinary decoding.
• Reproducibility. Because the code is said to be open, outside teams can check the numbers on their own Jetson Thor or DGX Spark hardware.

Nvidia is also pushing the software around this hardware. Our earlier report on CUDA 13.5 and Vera Rubin support covers the data center side of the same stack. VFlash is aimed at the other end, robots that cannot call a remote server for every decision.

Fan’s closing line is that the next inference revolution is “physical reasoning at the speed of reflex”. Whether VFlash delivers that outside Nvidia’s own setup will become clear once others reproduce the results.