The Hub is quiet on new model drops. Three blog posts carry the day. ESPnet points to a speech set that has been on the Hub since June. A community write-up shows a 1.11B pretrain fitting on a 6 GB laptop GPU. LeRobot shows a controller stack for humanoids built around latent motion tokens.
On the Hub
ESPnet YODAS v3: 1.1 million hours of open speech
ESPnet highlighted YODAS v3 on the Hub blog on Sep 27. The dataset card espnet/yodas3 was created on June 19 and is CC BY 3.0. The set is 1.1 million hours of multilingual speech at 48 kHz. Most files are stereo and timestamped. The blog lists more than 100 languages and 60 TB of OPUS audio. 1 2
1.11B LLM pretrain on a 6 GB laptop GPU
Rambarun Komaljeet posted a measured pretrain write-up on Sep 26. The run fits a 1.11B model on an RTX 4050 with 6 GB VRAM, peak 4.51 GB. The author uses BAdam, CPU offload, BitNet ternary weights and gradient checkpointing. Chinchilla scaling puts a full pretrain at about 375 days for 36B tokens, so fitting is not finishing. Code is Apache-2.0. 3
Bringing humanoids to LeRobot
The LeRobot blog from Sep 25 shows a whole-body controller for humanoids. π0.5 predicts 64-D SONIC latent motion tokens. A fast on-robot decoder turns them into joint targets. The stack runs 29-DoF motion on a Unitree G1. 4
