Tasks and inputs
Flora-Captioner is listed for Image-text-to-text, Video-text-to-text. It accepts video, text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
SyncDev model directory · Profile updated Nov 19, 2026
Flora-Captioner is NVIDIA's 27B non-commercial video model that writes one action instruction for a chosen interval of a robot manipulation video.
The specification score summarizes recorded features and limits. It is not a benchmark result.
Pricing conditions: Non-commercial research weights; no hosted API or price.
Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.
Accepts
Produces
Robot datasets are full of video and short on language. Flora-Captioner fills the gap one clip at a time. You hand it a target interval from a robot or human manipulation recording, optionally with the events just before and after, and it returns a single action instruction describing what happened inside the target.
NVIDIA released it on 15 October 2026 as one part of Flora, the data engine used to build GR00T-2. Internally it is called Flora-captioner-27B.
The base is Qwen3.5-27B. Training ran in two steps. Supervised fine-tuning updated the full language model and the visual merger without LoRA, while the patch embedding and vision transformer blocks stayed frozen. Reinforcement learning with Group Relative Policy Optimization followed.
This model is fussy about preprocessing, and the reference profile is worth following exactly:
| Setting | Reference value |
|---|---|
| Block order | BEFORE, TARGET, AFTER (omit missing context) |
| Target sampling | Duration-adaptive, nominal 8 Hz, 8–40 frames |
| Context sampling | Nominal 8 Hz, 4–12 frames per block |
| Frame budget | 64 frames total |
| Pixel caps | 262,144 per single-camera frame; 524,288 per three-camera mosaic |
| Token budget | 32,768 |
Three-camera input becomes one mosaic: main camera on top, left wrist bottom-left, right wrist bottom-right. Keep that layout fixed over time. Pass source timestamps to the processor and switch off its own frame sampling with do_sample_frames=False.
The target interval is half-open, [start_s, end_s), and BEFORE and AFTER are the neighbouring event intervals, not fixed padding. In the full Flora pipeline those intervals come from the companion segmenter.
Weights fall under the NVIDIA OneWay Noncommercial License. Research and development only; no commercial or production use. Outputs are annotations and must not drive a robot directly.
Sample two hundred segmented clips from your own robot, caption them, and have two annotators mark each instruction as correct, vague or wrong. Vague is the category to watch. A captioner that says "manipulate the object" is technically right and useless for training a language-conditioned policy. Compare against a general video-text model prompted the same way; the gain should show up mainly in specificity.
Siblings in the Flora engine are Flora-Rewarder, which scores success, and Flora-TempSeg, which finds the intervals. More NVIDIA robotics releases are on the provider page.
Compare the model on the work you need to complete.
Flora-Captioner is listed for Image-text-to-text, Video-text-to-text. It accepts video, text, image and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.
Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.
The weights are downloadable under NVIDIA OneWay Noncommercial License. Check runtime support and size the complete model and working context for your expected concurrency.
These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.
Tencent
WeMM-Video is Tencent's open Chinese-English omni-modal family (2B to 35B-A3B) that reasons over video frames and their soundtrack and answers in text.
NVIDIA
Flora-Rewarder is NVIDIA's 27B non-commercial reward model that scores robot episodes for success and grades each half-second of action quality.
perceptron
Perceptron Mk1.5 is an embodied reasoning model for robots, drones and glasses: video, image and audio in; text plus boxes, points and tracks out.
allenai
Molmo2-4B is Ai2's open 4B vision-language model for video, multi-image and image understanding, with pointing and object tracking built in.
Ai2 (Allen Institute for AI)
allenai/MolmoPoint-Vid-4B is a source-linked Hugging Face repository assigned to Video-text-to-text. Its recorded task interface accepts video, text and produces text.
chenjoya
chenjoya/videollm-online-8b-v1plus is a source-linked Hugging Face repository assigned to Video-text-to-text. Its recorded task interface accepts video, text and produces text.