NVIDIA’s GEAR robotics lab has released Flora, the set of models it uses to label the video that trains its GR00T robot foundation models. The lab estimates Flora labels an hour of robot video for $0.72, against $88.49 for GPT-6 Astra at maximum reasoning effort. The weights come under a noncommercial license, so for most enterprise teams the release is a price benchmark and a training recipe they cannot put into a production pipeline.
What NVIDIA released
Flora is three fine-tuned vision-language models. According to the project page and paper, dated October 20, a segmentation model built on Qwen3.5-9B splits a video into timestamped actions, a captioning model built on Qwen3.5-27B describes each action, and a reward model, Flora-Rewarder, also built on Qwen3.5-27B, grades how well a robot performed.
Those outputs map to three jobs in a robotics program. The paper says timestamped captions are used to filter and balance pretraining data, process rewards that score each step are used during reinforcement learning after pretraining, and outcome rewards that score a whole episode are used to monitor robots in deployment. Most existing tools, the paper says, cover only one of the three.
NVIDIA says Flora is already in production inside its own stack. The paper reports that it labels up to 200,000 hours of video per day across 512 GB300 GPUs. The three models are on Hugging Face, and the reward model’s card lists an October 15 release.
The cost comparison and its assumptions
The headline number comes from a fixed test of 50 episodes taken from DenseCaptionBench, NVIDIA’s own captioning benchmark: 20 human recordings, 15 parallel-gripper robot episodes and 15 dexterous-hand episodes, about 75 minutes of video in total. On that set the paper reports these quality scores and costs per hour of source video:
- Flora: 57.1% quality at an estimated $0.72
- GPT-6 Astra at maximum reasoning: 53.7% at $88.49
- Gemini 3.8 Flash: 38.3% at $8.48
- Claude Opus 5: 30.4% at $41.70
The two sides of that comparison are priced differently. Flora’s figure is an estimate for an eight-GPU GB300 deployment at an assumed rental price of $10 per GPU-hour. The API figures are what NVIDIA’s team was charged for its actual usage. A buyer with owned hardware, a different rental rate or lower utilization would get a different Flora number.
The paper also flags two limits on the quality scores. GPT-6 Astra served as the judge that graded every model’s captions, and because it was also one of the systems under test, the authors name generator and judge self-preference as a limitation. The paper also says the Flora and API runs were scored against reference snapshots that do not fully match.
Where frontier models still lead
The full benchmark table on the project page splits the result by task. Flora leads on human recordings and on reward labeling. On RoboRewardBench its mean absolute error is 40.0 against Astra’s 55.4, and on NVIDIA’s DenseRewardBench it catches 53.3% of failed steps inside an episode, where Astra catches none. The paper notes the cost of that sensitivity: Flora-Rewarder also marks about 11% of normal steps as failures.
Astra stays ahead on several tasks a robotics team would care about. On the robot-video subset of DenseCaptionBench it scores 71.9 on segmentation F1 against Flora’s 66.9, and 80.2 on caption quality against 73.7. On the public WGO-Bench, Astra’s gold-window accuracy is 77.4 against Flora’s 64.7. On NVIDIA’s RubricBench, where lower is better, Astra’s error is 3.6 and Flora’s is 6.5.
In our reading, the table supports a narrower claim than the price ratio suggests. A tuned open-weight model can match or beat the most expensive frontier API on some labeling tasks at a fraction of the cost, and the frontier model remains the better grader on others.
The license decides who can use it
NVIDIA’s paper describes the release as open-sourcing its data engine. The Hugging Face model card for Flora-Rewarder sets the terms: use is governed by the NVIDIA OneWay Noncommercial License, the model “is for research and development only”, and Flora “is not intended or supported for commercial or production use.” The card also says Flora outputs are annotations and should not serve as the sole safety monitor for a physical system.
The underlying Qwen3.5 base model is licensed under Apache 2.0, according to the same card. The paper documents the training method: supervised fine-tuning, followed by reinforcement learning for the segmentation and captioning models, on a mix of open-source datasets, licensed proprietary datasets and in-house data labeled by human annotators. A company could follow that method on the Apache-licensed base models with its own labeled data. It could not ship NVIDIA’s weights.
The weights sit on Hugging Face, the model hub covered in our earlier report NVIDIA Buys Hugging Face for $12.9 Billion, under NVIDIA’s own noncommercial terms.
What it means for the buying committee
Few enterprises train robot foundation models. More are collecting video from warehouses, plants, inspection rigs and pilot robots, and they buy labeling from service vendors or run frontier APIs over the footage. Flora gives those buyers three things to work with.
The first is a reference price. NVIDIA’s own estimate puts specialized open-weight labeling about 125 times below the most capable frontier API on its test set. That gap is wide enough to question any labeling contract or internal plan priced on frontier API rates, even after adjusting Flora’s figure for real hardware costs.
The second is a split between labeling tasks. Reward labeling and failure detection on long episodes are where Flora’s results are strongest. Fine-grained captions of robot video and rubric scoring are where Astra still wins. A data team can route each task to the cheapest model that meets its quality bar. Frontier API prices also move during a project, as with Gemini 4 Argon’s introductory pricing, which ends Nov. 30, so a long labeling program should not treat any API rate as fixed.
The third is a license check. Any team that wants Flora’s weights in a commercial pipeline needs terms from NVIDIA beyond the published noncommercial license. Procurement should get that answer before an engineering pilot builds on the models.
What to ask before you build on it
- Price your own GPU-hour. Rerun NVIDIA’s arithmetic with your actual hardware or rental rate and expected utilization before quoting the $0.72 figure internally.
- Test on your own footage. The 50-episode test set covers human, gripper and dexterous-hand video. Your cameras, tasks and failure modes will differ, and the judge in NVIDIA’s test was a competing model.
- Get the license answer in writing. Research use is covered. The model card excludes production use.
- Ask labeling vendors which models they run. If a vendor’s price assumes frontier API calls on every clip, Flora’s published numbers give you a basis to negotiate.
Source: NVIDIA Research
