NVIDIA Research and university collaborators posted VFlash on arXiv on 5 November 2026. It is a decoding method for vision-language models (VLMs: systems that take images and text together). The authors say it drafts complete visual plans, such as robot trajectories, in one forward pass rather than one coordinate token at a time.
On their MolmoSpaces manipulation pipeline, they measured end-to-end step latency falling from 3,552.7 ms to 703.3 ms on Jetson AGX Thor. On DGX Spark the same pipeline fell from 3,356.5 ms to 649.9 ms. Both comparisons use autoregressive decoding of the same 4B target. Those are author-run timings on NVIDIA hardware, not an independent check.
The paper’s pitch is engineering, not a new foundation model. Autoregressive VLMs spend most of a planning step on sequential coordinate tokens. VFlash freezes that target and trains a five-layer drafter (537 million trainable parameters) that predicts an entire primitive once its type and length are known.
Text still goes through verified speculative decoding. Visual primitives that pass a schema check are accepted without token-level agreement with the target. The project page is at NVIDIA’s LPR lab site.
What they measured on 3D trajectories
Open-loop 3D tests used InternA1, AgiBot and MolmoBot records with a shared primary-camera input. Against the same fine-tuned 4B autoregressive baseline, they report average displacement error (ADE, mean waypoint error in meters) falling by 18% to 49% for 3D-only targets. For joint 2D and 3D plans the reported ADE cut is 28% to 61%.
On AgiBot with the 2D+3D target, waypoint accuracy within 5 cm rises from 0.398 to 0.739 in their table. Those gains sit inside one lab’s stack. Readers who track how the big public leaderboards diverge will want an outside rerun before treating the percentages as firm.
Downstream, a world action model conditioned on VFlash plans scored 70.2% average success across four MolmoSpaces suites. Their baseline with the same action architecture scored 62.6%. Autoregressive plans in that setup scored 61.0%.
The largest suite lifts they report are Pick & Place (+11.6 points) and Close (+12.8). Driving heads on NAVSIM and Waymo look competitive with listed public systems after action-head reinforcement learning. Decoding speed is the clearer claim.
Hardware, baselines and what is missing
Latency tables break Thor, Spark and H100 into preprocess, vision, prefill, decoding and action. On Thor MolmoSpaces, decoding alone falls from 3,107.6 ms under autoregressive decoding to 258.2 ms with VFlash.
They also report about 25× faster single-rollout generation for dense 3D point-motion forecasting on a frozen MolmoMotion-AR target, with mixed results across HOT3D, WorldTrack and DAVIS. Stacks are customized TensorRT-Edge-LLM builds in FP16.
That is careful work, and it is still NVIDIA measuring NVIDIA silicon. The same caution applies when AMD discloses Helios racks without breaking out GPU revenue: platform numbers matter, and so does who ran them.
The paper does not give an independent timing on non-NVIDIA edge silicon. It also omits a public open-weight release of the full planning VLM used in most tables, and any head-to-head against outside production robot stacks.
Box generation needs an AR fallback to hold mean F1 near the baseline. Without it, Dense200 mean F1 drops roughly 20 points. Treat the roughly 80% latency cut and up-to-61% ADE cut as author claims tied to named datasets, GPUs and a named baseline decoder.
What to do this week
If you ship VLM planners on Jetson-class hardware, read the latency appendix against your token counts and action-chunk schedule. Ask vendors for timings that name the baseline decoder, batch size and whether action-head feature extraction counts.
We later looked at AIRoA-0’s author-run bimanual success rates from handheld demos, another case where the protocol is disclosed but outside replication is still missing.
If you evaluate research claims, demand the ADE table with the matching AR row, not a relative percentage alone. The arXiv abstract and PDF are primary. Secondary write-ups such as Interesting Engineering’s summary mostly restate their numbers.
Sources
Get the weekly email
The week's most useful AI news, a tutorial and an interview.
Lars Strother
Lars Strother is the senior editor for models and hardware at AI Observer, covering models and research, chips and hardware, and robotics. He came to journalism from engineering and still reads the benchmark tables before the press release. His coverage is measured and skeptical, and he wants to know what was tested, on what, and against which baseline.
Spotted a mistake? Tell us on the corrections page.