Flora
Nvidia open-sources Flora Flora consists of fine-tuned Qwen 3.5 models Flora employs a 9B Qwen 3.5 model to split videos into action segments Flora employs a 27B Qwen 3.5 model to describe actions and a second 27B model as Flora-Rewarder Flora employs the Flora-Rewarder model for evaluating episode success and step quality Flora runs during the training of Nvidia's GR00T models Flora models are available on Hugging Face Flora leads on action boundary detection in the WGO-Bench benchmark Flora leads across all metrics on DenseCaptionBench for human recordings Flora ranks second behind Astra on DenseCaptionBench for robot videos· Graphiti 生成的实体摘要
关系变化
按生效时间- FloraAVAILABLE ONHugging Face
Flora models are available on Hugging Face
- FloraHAS MODEL27B model
Flora employs a 27B Qwen 3.5 model to describe actions and a second 27B model as Flora-Rewarder
- FloraLEADS INWGO-Bench
Flora leads on action boundary detection in the WGO-Bench benchmark
- FloraLEADS INDenseCaptionBench
Flora leads across all metrics on DenseCaptionBench for human recordings
- FloraUSED IN TRAINING OFGR00T
Flora runs during the training of Nvidia's GR00T models
- NVIDIAOPEN SOURCESFlora
Nvidia open-sources Flora
- FloraHAS MODELQwen3.5-9B
Flora employs a 9B Qwen 3.5 model to split videos into action segments
- FloraBUILT ONQwen3.5
Flora consists of fine-tuned Qwen 3.5 models
- FloraHAS MODELFlora-Rewarder
Flora employs the Flora-Rewarder model for evaluating episode success and step quality
- FloraSECOND PLACE INDenseCaptionBench
Flora ranks second behind Astra on DenseCaptionBench for robot videos
关系是 Graphiti 从报道中提取的事实陈述，不是原文引述；失效时间表示后来的报道显示这条关系已不成立。