Show HN: 27B open model to grade robot data 48x cheaper/27x faster vs. 6 Astra

Josef Chen released RoboJev, a 27B-parameter open-weights model (JEV-27B-VL, Apache-2.0) that grades robot task success from video, claiming it is 48x cheaper and 28x faster than GPT-6 Astra. Grading 68,555 episodes from 7,641 public LeRobot eval datasets on Hugging Face — 501 hours of video — costs $13.16 of GPU time, versus the $637 GPT-6 Astra would charge, according to Chen. The model was fine-tuned with LoRA on 52,598 robot episodes, and Chen released the weights, all three evaluation sets, a LeRobot reward plugin and a live-verdict video overlay.

Josef Chen on X: "RoboJev: people are watching thousands of videos to label task success. Now, you can grade robot data faster than you make it. 48x cheaper and 28x faster than 6 Astra. 27B parameters, open weights." RoboJev: people are watching thousands of videos to label task success. Now, you can grade robot data faster than you make it. 48x cheaper and 28x faster than 6 Astra. 27B parameters, open weights. RoboJev: people are watching thousands of videos to label task success. Now, you can grade robot data faster than you make it. 48x cheaper and 28x faster than 6 Astra. 27B parameters, open weights. 3/ Grading 68,555 episodes from 7,641 public LeRobot eval datasets on Hugging Face, 501 hours of video, takes $13.16 of GPU time. GPT-6 Astra would charge $637. 5/ How it's built: JEV-27B-VL, fine-tuned with LoRA on 52,598 robot episodes. Open: weights Apache-2.0 , all three evaluation sets, the LeRobot reward plugin and the live-verdict overlay from the video. Details and demo: Top: the ring lands and RoboJev goes to 1.00. Bottom: the ring sits on the peg for a while, then falls off, and the score drops from 0.82 to 0.04. Try it on your own robot video: huggingface.co/spaces/josefch…