🔍 Read the full analysis: The Promise And License Catch Of Nvidia Flora’s 125× Cheaper Video Labeling on ThorstenMeyerAI.com
Get hardware and tech essentials delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Nvidia’s GEAR lab released Flora, three models for segmenting, captioning and scoring robot video. Nvidia estimates captioning costs $0.72 per video hour versus $88.49 for GPT-6 Astra, but the first figure is modeled, the comparison uses a limited test set, and Flora’s license allows research and evaluation—not commercial use.
Nvidia’s GEAR robotics lab has released Flora, three models that segment, caption and score robot video, alongside a claim that its captioning can cost about 125 times less than GPT-6 Astra. The comparison is not a measured, like-for-like production test: Nvidia models Flora’s cost at $0.72 per video hour, while Astra’s $88.49 figure comes from API usage on a 50-episode test set. Flora’s license also limits use to research or evaluation, ruling out commercial deployment under its stated terms.
The Flora release consists of three models built on Alibaba’s Qwen3.5: a 9-billion-parameter temporal segmenter that identifies action intervals, a 27-billion-parameter captioner that describes each interval, and a 27-billion-parameter rewarder that scores a robot’s actions. Nvidia says the models were used as a data engine for its GR00T-2 robot system and were trained on fewer than 10,000 hours of human and robot video. The models and inference code are available through Hugging Face repositories.
Nvidia’s paper says the cost comparison used 50 episodes from its DenseCaptionBench, totaling 4,483 seconds, or about 75 minutes, of video. GPT-6 Astra’s reported cost was $88.49 per video hour at public API rates dated October 4, 2026. Flora’s $0.72 estimate assumes eight GB300 GPUs rented at $10 per GPU-hour and modeled throughput; Nvidia gives a range of $0.51 to $1.26, excluding storage and other overhead. Dividing the two headline figures gives a ratio of about 122.5 to one, rounded by Nvidia to approximately 125×.
The paper reports a mixed quality result on that test: Flora scored 57.1%, Astra 53.7%, Gemini 3.8 Flash 38.3% and Claude Opus 5 30.4%. Across other reported benchmarks, Flora performed better on human-video examples, action boundaries and partial-success scoring. Astra performed better on robot videos, caption accuracy and rubric scoring. The source account says most tests were run by Nvidia, and the reported cost claim concerns captioning; the reward model is not included in the 125× calculation.
Infographic 01 · What Flora does
Three models that turn robot video into training labels.
Flora cuts a recording into actions, describes each one and grades how well it went. The 200,000-hour figure is Nvidia’s own production claim.
Flora-TempSeg
9B · adapter on Qwen3.5-9B
- Job: Finds where each action starts and ends
- In: Video windows (one camera, or main + two wrist cameras)
- Out: Time intervals as JSON
Flora-Captioner
27B · fine-tuned Qwen3.5-27B
- Job: Describes each action, using the video before and after
- In: One interval plus context
- Out: One English instruction, e.g. “Place the blue mug on the tray with your right hand.”
Flora-Rewarder
27B · fine-tuned Qwen3.5-27B
- Job: Judges how well the robot did
- In: An episode or a 16-second window, plus the task
- Out: Score 1–5 per episode · normal / suboptimal / failure per 0.5 s · rubric scores
Three labels, three uses
Why robot models need each one, per the paper
| Label | Used for | Made by |
|---|---|---|
| Timestamped captions | Filtering and rebalancing pretraining data | TempSeg + Captioner |
| Process rewards | Reinforcement learning after pretraining: which step went wrong | Rewarder |
| Outcome rewards | Monitoring robots after deployment | Rewarder |
What it is not
model cards
- Not a robot controller: outputs are annotations only
- Not meant to be the sole safety monitor for a physical system
- Not for commercial or production use (noncommercial license)
- Hardware: tested on Nvidia H100, Linux; ~55 GB of weights per 27B model (estimate)
Sources: Flora paper, Nvidia (20 Oct 2026); Hugging Face model cards (15 Oct 2026); Nvidia GEAR project page. Weight-size estimate: Thorsten Meyer (27B parameters × 2 bytes).
Powered by Thorsten Meyer AI
Lower Labeling Costs, With Limits
Video labels help robotics teams identify what happened in recordings, where each action began and whether an attempt succeeded. Reducing the cost of producing those labels could make it easier to process larger training datasets, a potentially important step for teams developing robot foundation models. Nvidia says Flora can process up to 200,000 hours of video per day on 512 GB300 GPUs; that is a company-reported capacity figure, not independent evidence of routine production throughput.
The headline saving should not be treated as a guaranteed operating cost. One side of the comparison is an assumed GPU rental budget; the other is API usage on a limited benchmark. Nvidia’s paper says the production GPU-hour denominator has not been verified. Even if the estimate points to a substantial cost reduction, companies would need to account for their own hardware, utilization, storage, engineering and quality requirements before judging the economics.
The license is a separate constraint. Flora is released under the NVIDIA OneWay Noncommercial License for research or evaluation, and derivatives inherit the restriction, according to the source material. That makes the models available for study and testing, but not a ready-made free commercial labeling service. The models are also large: the source estimates that each 27-billion-parameter model needs roughly 55 GB for weights at 16-bit precision, placing local use beyond typical gaming PCs.
Products Worth Considering
Apple MacBook Air 13-inch with M5 Chip, 16GB Memory, 512GB SSD, 12MP Camera, Wi-Fi 7, Midnight
Apple MacBook Pro 14.2-inch with M3 Max, 14-Core CPU, 30-Core GPU, 36GB RAM, 1TB SSD, Space Black
Apple MacBook Air 15-inch Laptop with M5 Chip, 16GB Memory, 512GB SSD, and Liquid Retina Display
As an affiliate, we earn on qualifying purchases.
How Flora Labels Robot Video
Robot-learning systems use recordings of actions such as placing an object or folding fabric. Those recordings need labels that describe actions and connect them to timestamps. The Flora paper describes three related tasks: timestamped captions to identify actions, step-level rewards to mark normal, suboptimal or failed behavior, and episode-level scores to assess an entire attempt. These labels can help researchers organize training data and evaluate robot performance.
Flora’s segmenter proposes the time intervals, its captioner writes an instruction for each interval using surrounding video for context, and its rewarder can grade short steps or score an episode against a rubric. Nvidia’s comparison with Astra focused on the captioning benchmark, not the full range of tasks. The models’ release therefore provides a possible alternative for research workflows, but the reported 125× figure should not be read as a measured saving across every labeling task.
The benchmark included 20 human-hand episodes, 15 with two-finger grippers and 15 with dexterous robot hands. Results varied by category and evaluation measure. The supplied source reports Flora’s advantages on some human-video and partial-success tests, while Astra retained advantages on robot-video captioning and rubric-based scoring. Those differences matter because a model’s usefulness depends on the type of footage and labels a robotics team needs.
“The production GPU-hour denominator has not been verified.”
— Nvidia, in the Flora paper
Cost and Quality Still Need Testing
The available figures do not establish what Flora would cost or how it would perform across a company’s own production data. Nvidia’s estimate depends on an assumed $10 per GPU-hour rate and modeled throughput, and excludes storage and overhead. The test covered about 75 minutes of video selected for Nvidia’s benchmark; the source does not provide an independent production-scale replication. It also reports 106 of 7,062 Astra attempts with no usage data, making the API cost a known-usage subtotal.
Quality is not settled by the overall benchmark score. Flora’s reported lead on the test set coexists with areas where Astra scored better, including robot-video captions and rubric scoring. The supplied source does not give enough detail to establish how these results would generalize to other camera setups, robot tasks or data distributions. Nor does the 125× comparison include all three Flora models or the cost of the full labeling pipeline.
Commercial rights and practical deployment details also remain constraints. The stated license is noncommercial, while the large models call for data-center-class hardware according to the source estimate. Readers should treat the performance, throughput and cost figures as reported results and estimates, not as an independent guarantee of commercial readiness.
Independent Tests and License Terms
The next useful evidence would be independent replication of Flora’s throughput and quality on larger, varied video sets, with all GPU, storage and operating costs stated. Comparisons would also be clearer if they used equivalent hardware assumptions and measured the same labeling tasks, rather than comparing modeled GPU costs with API charges.
Organizations considering the release will need to review the NVIDIA OneWay Noncommercial License and assess whether their planned work qualifies as research or evaluation. Any commercial deployment would require terms that permit it. Further reporting should track whether Nvidia publishes production measurements, additional benchmark results or revised licensing options; the supplied material does not identify a date for such updates.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab: one identifies action segments, one captions them, and one scores robot behavior.
Is Flora really 125 times cheaper than GPT-6 Astra?
Nvidia’s estimate compares a modeled Flora cost of $0.72 per video hour with an Astra API cost of $88.49 on a 50-episode test. The ratio is about 122.5, rounded to approximately 125×, but the two figures are not measured on the same cost basis.
Can businesses use Flora commercially?
The source says the models use the NVIDIA OneWay Noncommercial License, limited to research or evaluation purposes, with the restriction applying to derivatives. Businesses should review the license terms rather than assume the release permits commercial use.
Does Flora outperform GPT-6 Astra on every test?
No. The reported results are mixed: Flora led on some human-video, action-boundary and partial-success measures, while Astra performed better on robot videos, caption accuracy and rubric scoring.
What remains unverified about Nvidia’s claim?
Nvidia describes Flora’s cost as a model based on an assumed GPU rental rate and throughput, and says the production GPU-hour denominator has not been verified. The benchmark is limited, and the 125× figure covers captioning, not every Flora model or the full cost of deployment.
Source: ThorstenMeyerAI.com
