🔍 Read the full analysis: Nvidia Flora: The 125× Robot Video Labeling Savings And The Fine Print on ThorstenMeyerAI.com
Get business pricing on monitors, keyboards and dev gear
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
TL;DR
Nvidia’s GEAR lab released Flora, three models for segmenting, captioning and scoring robot videos. Nvidia estimates captioning costs $0.72 per video hour versus $88.49 for GPT-6 Astra, a 122.5× ratio, but Flora’s figure is modeled, its benchmark results are mixed, and its license restricts use to research or evaluation.
Nvidia’s GEAR robotics lab has released Flora, a set of three models for labeling robot videos, and estimates that its captioning system can process video for $0.72 per hour, compared with $88.49 for GPT-6 Astra. The claimed approximately 125× cost advantage is based on a modeled GPU budget compared with measured API usage, not two directly measured production costs; Flora’s license also limits use to research or evaluation.
Flora combines three models built on Alibaba’s Qwen3.5 family. The 9-billion-parameter Flora-TempSeg identifies the time intervals in which actions occur. A 27-billion-parameter Flora-Captioner writes an instruction for each interval, while the 27-billion-parameter Flora-Rewarder scores an entire episode, grades short steps as normal, suboptimal or failure, or evaluates them against a user-provided rubric. The systems are intended to help label video used to train and monitor robots.
The paper’s cost test used 50 episodes totaling 4,483 seconds, or about 75 minutes, from Nvidia’s DenseCaptionBench. GPT-6 Astra was charged using public API rates from Oct. 4, 2026; Nvidia reports a cost of $88.49 per video hour. Flora’s $0.72 figure was calculated using eight GB300 GPUs, an assumed price of $10 per GPU-hour and modeled throughput. Nvidia gives a range of $0.51 to $1.26 per video hour, excluding storage and other overhead. The paper says the production GPU-hour denominator has not been verified.
Dividing the reported Astra cost by Flora’s modeled figure produces a 122.5× ratio, rounded in Nvidia’s presentation to approximately 125×. The calculation covers captioning, not the reward model, and the source reports 106 attempts with no usage data in the API test. Nvidia also reports an overall benchmark score of 57.1% for Flora, compared with 53.7% for Astra, 38.3% for Gemini 3.8 Flash and 30.4% for Claude Opus 5. Those results come from the same 50 Nvidia-selected episodes, and the paper’s broader tests show that performance varies by task.
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 Costs Could Expand Robot Training
Video labeling is a labor-intensive step in building robot datasets. Timestamped captions help researchers identify and balance actions in large collections, while step-level and episode-level scores can be used to assess performance and support reinforcement learning. If Flora’s cost estimate holds at larger volumes, automated labeling could reduce one expense involved in preparing data for robot training.
The result does not remove other constraints. The models still need suitable video, computing hardware and evaluation. The 27-billion-parameter models are designed for data-center-class hardware; the source estimates that each would require roughly 55 GB for weights at 16-bit precision. And businesses cannot treat the released models as a free commercial labeling service under the stated license. For now, Flora’s reported efficiency is a research result and potential technical reference, not evidence that companies can deploy it commercially at the quoted price.
Products Worth Considering
Jackery Solar Generator 1000 V2 with 200W Solar Panel
Apple iPad Air (2025) 13‑inch Wi‑Fi 256 GB – Space Gray (M3 chip)
DJI Osmo Mobile 8 Standard Combo Phone Gimbal Stabilizer
As an affiliate, we earn on qualifying purchases.
Three Models, Three Labeling Tasks
Robot-learning systems are trained on recordings of actions, such as a gripper placing a mug or a person folding a towel. Those recordings need labels that describe what happened and when, as well as indications of whether a step or full attempt succeeded. The Flora paper presents its models as a way to produce those different labels from video rather than relying on a single captioning tool.
Nvidia says Flora was used as a data engine for its GR00T-2 robot models and was trained on fewer than 10,000 hours of human and robot video. The models and inference code are available through Hugging Face repositories, according to the source material. All three use the NVIDIA OneWay Noncommercial License, which permits research or evaluation and extends the restriction to derivatives. Nvidia itself may use the models commercially, but the terms do not grant that permission to other businesses.
“research or evaluation purposes only”
— NVIDIA OneWay Noncommercial License
Cost and Quality Still Need Testing
The comparison does not establish the cost of running Flora in a commercial production environment. Nvidia’s Flora figure assumes a GPU rental price and modeled throughput, while Astra’s figure comes from reported API usage and public rates. The test covers roughly 75 minutes of video, and the source notes that some API attempts had no reported usage. Nvidia says storage and overhead are excluded and that the production GPU-hour denominator is unverified.
Quality results also depend on the task and video type. Flora scored better on the reported overall test and on human-video examples, action boundaries and partial successes. But the source says Astra performed better on robot videos, caption accuracy and rubric scoring. Most tests were run by Nvidia, and the provided material does not establish how results would generalize to different datasets, hardware configurations or longer production workloads. The license’s application to specific business arrangements may also require review of the full terms.
More Evidence Needed at Scale
The next test is whether Flora’s modeled throughput and cost hold when the models process larger, varied datasets on real infrastructure. Independent evaluations could help clarify how consistently the systems label human and robot footage, how much review their outputs require and whether the benchmark differences persist beyond Nvidia’s test set.
Organizations considering the models will also need to check the license before using them in a product or commercial data pipeline. Nvidia’s release establishes an available research tool and a documented cost estimate; it does not yet establish a verified production price or grant outside companies commercial rights. The source material does not specify a date for further benchmarks, revised cost measurements or a change to the licensing terms.
Key Questions
What is Nvidia Flora?
Flora is a set of three video models from Nvidia’s GEAR robotics lab: one segments actions by time, one captions those segments, and one scores robot performance.
Is Flora really 125 times cheaper than GPT-6 Astra?
That is Nvidia’s rounded comparison, not a like-for-like measurement of two production bills. Astra’s reported $88.49 per video hour is based on API usage, while Flora’s $0.72 is a modeled GPU cost using an assumed $10 per GPU-hour. The division is 122.5×, rounded to approximately 125×.
Does the cost comparison cover all three Flora models?
No. The stated cost comparison covers captioning. It does not include the reward model, and the source says storage and other overhead are excluded from Flora’s estimate.
Can companies use Flora commercially?
The models are released under the NVIDIA OneWay Noncommercial License, described as allowing research or evaluation only. The source says derivatives inherit that limit; businesses should review the full license before use.
Where did Flora perform better or worse?
In the reported tests, Flora scored well on human videos, action boundaries and partial-success cases. The source says GPT-6 Astra remained ahead on robot videos, caption accuracy and rubric scoring, so Flora’s results are not uniformly superior.
Source: ThorstenMeyerAI.com
