ModelsPublished All news from that day
NVIDIA says its Flora data engine labels 200,000 hours of video a day on 512 GB300 GPUs.
NVIDIA Research's page says Flora, used to train its GR00T robot models, runs on 512 GB300 GPUs and costs roughly a 125th of GPT-6 Astra to run; the Hugging Face card limits it to research.
Check our sources · 8 facts from 2 sourcesNVIDIA Research describes Flora as a data engine used for training its GR00T robot models, producing captions with timestamps and reward labels for robot and human video. NVIDIA says it labels 200,000 hours of data per day across 512 GB300 GPUs, and that it is roughly 125 times cheaper than GPT-6 Astra.
NVIDIA's own benchmark table shows Flora ahead of GPT-6 Astra on some tests and behind on others.
The Flora-TempSeg model card on Hugging Face lists a release date of 15 October 2026 and a 9-billion-parameter segmenter post-trained from Qwen3.5-9B. The card says the model is governed by the NVIDIA OneWay Noncommercial License and is not intended or supported for commercial or production use.
Your reaction
We count reactions per story and day, never who reacted. The counts help us choose what goes in the monthly issue. If you are signed in, your own page shows yours too.
We count reactions per story and day, never who reacted. They help choose what goes in the monthly issue.
Check our sources
Every sentence above is checked against these 2 sources.
1 Flora: a data engine for labeling human and robot video
Open the source archived copy-
NVIDIA Research's Flora page describes Flora as "an industrial-strength data engine used for training GR00T models".
citean industrial-strength data engine used for training GR00T models
-
NVIDIA Research's Flora page says Flora is "battle tested at scale, labeling 200,000 hours of data per day across 512 GB300s".
citelabeling 200,000 hours of data per day across 512 GB300s
-
NVIDIA Research's Flora page says the model is "substantially cheaper than API-based alternatives, roughly ~125× cheaper than GPT-6 Astra". The comparison is NVIDIA's own and the page gives no per-hour price.
citeroughly ~125× cheaper than GPT-6 Astra
-
NVIDIA Research's Flora benchmark table lists Flora at 40.0 and GPT-6 Astra at 55.4 on RoboReward MAE (lower is better), and Flora at 64.7 against GPT-6 Astra's 77.4 on WGO-Bench gold-window accuracy (higher is better). The benchmarks are NVIDIA's own.
citeRoboReward MAE 40.0 vs 55.4; WGO-Bench gold-window accuracy 64.7 vs 77.4 (Flora vs GPT-6 Astra)
-
NVIDIA Research's Flora page labels its benchmark results groups "Timestamped captions" and "Reward annotations", for human and robot video.
Toned down to what the source says
citeTimestamped captions and Reward annotations appear as benchmark group labels in the results table; the page's section headings are Timestamped Action Instructions and Reward Modeling.
2 nvidia/Flora-TempSeg model card
Open the source archived copy-
The Flora-TempSeg model card on Hugging Face gives the release date as 10/15/2026 and describes Flora-segmenter-9B as having 9.0 x 10^9 parameters, developed from Qwen3.5-9B.
cite10/15/2026; Flora-segmenter-9B: 9.0 x 10^9; This model was developed based on Qwen3.5-9B.
-
The model card says: "Use of this model is governed by the NVIDIA OneWay Noncommercial License (22 March 2022)."
citeUse of this model is governed by the NVIDIA OneWay Noncommercial License (22 March 2022).
-
The model card says: "Flora is not intended or supported for commercial or production use."
citeFlora is not intended or supported for commercial or production use.
Every sentence above is checked against these 2 sources. Each fact has its own address you can cite.
Source 1 · NVIDIA Research (GEAR lab)
Flora: a data engine for labeling human and robot video
NVIDIA Research's Flora page describes Flora as "an industrial-strength data engine used for training GR00T models".
an industrial-strength data engine used for training GR00T models
NVIDIA Research's Flora page says Flora is "battle tested at scale, labeling 200,000 hours of data per day across 512 GB300s".
labeling 200,000 hours of data per day across 512 GB300s
NVIDIA Research's Flora page says the model is "substantially cheaper than API-based alternatives, roughly ~125× cheaper than GPT-6 Astra". The comparison is NVIDIA's own and the page gives no per-hour price.
roughly ~125× cheaper than GPT-6 Astra
Show all 5 factsShow fewer facts
NVIDIA Research's Flora benchmark table lists Flora at 40.0 and GPT-6 Astra at 55.4 on RoboReward MAE (lower is better), and Flora at 64.7 against GPT-6 Astra's 77.4 on WGO-Bench gold-window accuracy (higher is better). The benchmarks are NVIDIA's own.
RoboReward MAE 40.0 vs 55.4; WGO-Bench gold-window accuracy 64.7 vs 77.4 (Flora vs GPT-6 Astra)
NVIDIA Research's Flora page labels its benchmark results groups "Timestamped captions" and "Reward annotations", for human and robot video.
Toned down to what the source saysTimestamped captions and Reward annotations appear as benchmark group labels in the results table; the page's section headings are Timestamped Action Instructions and Reward Modeling.
Source 2 · NVIDIA on Hugging Face · 15 Oct 2026
nvidia/Flora-TempSeg model card
The Flora-TempSeg model card on Hugging Face gives the release date as 10/15/2026 and describes Flora-segmenter-9B as having 9.0 x 10^9 parameters, developed from Qwen3.5-9B.
10/15/2026; Flora-segmenter-9B: 9.0 x 10^9; This model was developed based on Qwen3.5-9B.
The model card says: "Use of this model is governed by the NVIDIA OneWay Noncommercial License (22 March 2022)."
Use of this model is governed by the NVIDIA OneWay Noncommercial License (22 March 2022).
The model card says: "Flora is not intended or supported for commercial or production use."
Flora is not intended or supported for commercial or production use.
Topics
The morning email
On the mornings we publish: the three top stories and up to four short ones. Free.
We email you a link to confirm. An issue may include one sponsor, always labelled Sponsored · Advertisement. Our emails count opens and clicks, not who made them. Unsubscribe in one click. What we keep