Futureweb

Robotics

Flora labels are not a controller

NVIDIA's Flora paper is dated October 20. Model cards are October 15. The weights are research-only, and the 125-times cost line is the paper's.

The short version

NVIDIA published Flora labeling models, with a paper dated October 20 and model cards dated October 15. The license is research and noncommercial. The paper says the models are not a robot controller or a sole safety monitor.

A paper, then a later write-up

NVIDIA's Flora paper, from the GEAR lab, is dated October 20, 2026. The Hugging Face cards for Flora-TempSeg, Flora-Captioner, and Flora-Rewarder are dated October 15. The Decoder wrote on October 25 that NVIDIA had open-sourced Flora to make robot-video labeling cheaper. The release is mid-October. October 25 is the day an outlet wrote about it. The models label video for robot data. The paper says they are not a robot controller and not a sole safety monitor. The document named in the sources is the record this page will keep.A lamp proposal is not a controller either.

Whose numbers

The paper's own cost line is about 125 times cheaper than GPT-6 Astra, and a throughput line of up to 200,000 hours a day across 512 GB300 chips. Those rates are NVIDIA's. The Decoder repeats them. This page did not measure the bill. Flora-TempSeg is described as 9 billion parameters on a Qwen3.5-9B base. Captioner and Rewarder are described near 27 to 28 billion. The paper says the stack was used for GR00T data. That use is the paper's account of NVIDIA's own training pipeline.A round is not a fleet.

Two benchmarks, not one winner

On the paper's DenseRewardBench, Flora-Rewarder's failure recall is 53.3 percent against baselines under 11.7 percent, and about 11 percent of normal steps are labeled as failures. That is the paper's table, not an independent test. On RubricBench the paper lists Flora-Rewarder's pooled L1 at 6.5 and Astra's at 3.6. Lower is better on that error, so Astra leads it, while the paper says 6.5 still beats nine of the ten general-purpose models. A headline that says Flora beats Astra has to say which table. This page will not collapse them.On-device inference is a different limit.

The license is the scope

The weights carry NVIDIA's OneWay Noncommercial License of 22 March 2022. The Qwen base models are Apache 2.0. Research and noncommercial use is what the weight license allows. A commercial robot product, a production line, or a safety system is outside that grant. The paper's own warning matches the license: not a controller, not the only safety monitor. A reader who wants to ship a product needs a different agreement than the card this page read. Asked directly, Does a lower RubricBench error mean Flora beat Astra on every test? No. The paper's RubricBench pooled L1 is 6.5 for Flora-Rewarder and 3.6 for Astra, and lower is better, so Astra leads that table. DenseRewardBench is a different result.The key stays with the holder.

Who the labels are for

The users in scope are research groups who accept the noncommercial terms and want labels on robot video. A factory that wants a certified safety stop is not in scope. GR00T, as the paper describes it, is NVIDIA's own use of the labels. It is not a promise that every GR00T robot in a lab is running Flora as its controller. Asked directly, Is Flora 125 times cheaper than GPT-6 Astra? That multiple is the paper's claim. This page did not rerun the cost comparison.

What the Decoder adds

The Decoder is the outside account: it dates the news to the week it published and leads with the 125-times claim. Both of those choices are the outlet's framing of a paper that already existed. The useful correction is the calendar. Cards on the 15th, paper on the 20th, newspaper on the 25th. Asked directly, Can a company ship Flora in a commercial robot? The NVIDIA OneWay Noncommercial License, dated 22 March 2022, is research and noncommercial. The Qwen bases are Apache 2.0. The Flora weights are not a commercial product license.

What this page will not say

It will not say the weights are a commercial open-source release, that October 25 is the ship date, or that Flora won a single comparison against Astra. It will say NVIDIA posted research labeling models under a noncommercial license, with the paper's own cost and benchmark figures labeled as the paper's. Asked directly, What are the three Flora models? Flora-TempSeg, a 9 billion parameter model on a Qwen3.5-9B base; Flora-Captioner; and Flora-Rewarder, both described around 27 billion parameters. The document named in the sources is the record this page will keep.The advisory names the license.

The open item

Open items are an independent rerun of the cost claim and a commercial license, if NVIDIA ever offers one. Neither is in the October documents. The documents settle the model names, the dates, the noncommercial grant, and the paper's warning about control and safety. The short answer this page will stand on is this: NVIDIA published Flora labeling models, with a paper dated October 20 and model cards dated October 15. The license is research and noncommercial. The paper says the models are not a robot controller or a sole safety monitor.

A partnership is a different brief

Teradyne and Sereact's October 21 pairing is a product announcement about arms already in warehouses. Flora is a labeling stack. They share a desk and not a dependency. This page does not treat a model card as a deployment, and it does not treat a press release as a benchmark. Asked directly, When were the Flora weights posted? The Hugging Face cards this page read are dated October 15, 2026. The paper is dated October 20. The Decoder's October 25 story is an account, not the release date.

Sources

The reports this brief is filing. Futureweb did not republish them.

  1. Flora paper
  2. Flora-TempSeg model card
  3. The Decoder

Questions

When were the Flora weights posted?

The Hugging Face cards this page read are dated October 15, 2026. The paper is dated October 20. The Decoder's October 25 story is an account, not the release date.

What are the three Flora models?

Flora-TempSeg, a 9 billion parameter model on a Qwen3.5-9B base; Flora-Captioner; and Flora-Rewarder, both described around 27 billion parameters.

Can a company ship Flora in a commercial robot?

The NVIDIA OneWay Noncommercial License, dated 22 March 2022, is research and noncommercial. The Qwen bases are Apache 2.0. The Flora weights are not a commercial product license.

Is Flora 125 times cheaper than GPT-6 Astra?

That multiple is the paper's claim. This page did not rerun the cost comparison.

Does a lower RubricBench error mean Flora beat Astra on every test?

No. The paper's RubricBench pooled L1 is 6.5 for Flora-Rewarder and 3.6 for Astra, and lower is better, so Astra leads that table. DenseRewardBench is a different result.