Hugging Face launches Real World VoiceEQ benchmark for voice AI quality

A robotics demo shows a worker holding a bubble tea cup, with fingers wrapping its far side and visibility changing between camera views, complicating reconstructed hand overlays.

U.S. District Judge Edward Davila gave OpenAI five days to produce already-collected interview materials and 14 days for the rest, covering interviews with ex-Apple staff who applied to its hardware and robotics groups. Apple also won access to private LINE messages from people named in the case.

FastH3-Omni arrives October 20 with open weights and local Spark/RTX/MLX support, per r/StableDiffusion. It is reference-conditioned, supporting character and setting reference images, first- and last-frame conditioning, and up to 3 image references in real-time mode.

Nvidia uploaded a GR00T-N1.7 checkpoint to Hugging Face, tagged AppleToPlate-cosine40k-checkpoint-15000, suggesting a robot manipulation policy trained on an apple-to-plate task for 40k steps with a cosine schedule.

Both Microsoft image models are now callable via AI Gateway with one API key and no provider account, supporting text-to-image and edits with up to five reference images at up to 1536 × 1536 output. MAI-Image-2.6 targets final assets; the Flash variant trades quality for faster, cheaper iteration.

Kai Van Brunt, an MIT '25 graduate in Sara Beery's BeeryLab, built Molmo2Fish on Ai2's open Molmo 2 vision-language model to follow salmon in underwater sonar video. The prototype revises its own tracking predictions when a researcher describes the error in everyday language.