The Hub is leaning into humanoids, typed decisions, and all-day speech. A community post on September 25 shows whole-body control of a Unitree G1 inside LeRobot. Laya is a small Apache-2.0 model that answers typed questions with a probability instead of writing a paragraph. Edge0 posted Audio8-ASR-Infinite, a streaming recognizer meant to run all day without the usual long-form drift.
On the Hub
LeRobot × Unitree G1
Martino Russi, Steven Palma, and Pepijn Kooijmans posted a Hugging Face community article on September 25 about bringing humanoids into LeRobot.1 A π0.5 policy predicts SONIC motion tokens from language, camera, and robot state. A fast controller on the Unitree G1 then turns those tokens into stable 29-DoF whole-body motion, because a biped has to stay balanced and a vision-language-action model is too slow to drive the joints directly. Printable wrist adapters, OpenArm grippers, and a backpack enclosure live in a hardware-mods dataset that has been on the Hub since late August. The OpenHLM recipe they follow saturates around 40 demonstrations per task.
Laya
Laya has been on the Hub since September 18 under Apache 2.0.2 It is a non-autoregressive System-1 decision model: you hand it a state (an email, a ticket, a JSON blob) plus typed questions, and it returns typed answers with calibrated probabilities in one forward pass, about 33 milliseconds. It never generates text, so there is nothing to parse. The English checkpoint is a 421 million parameter ModernBERT-large, and a multilingual twin routes non-English text automatically.
Audio8-ASR-Infinite
Edge0 posted Audio8-ASR-Infinite on the Hub on September 21 as an Apache-2.0 preview.3 It is native streaming speech recognition with a selectable 80, 120, or 160 millisecond audio clock and a transcription delay you can set between 240 and 560 milliseconds. A rolling KV cache is meant to keep memory and latency constant for unbounded, all-day transcription, using an adapted vLLM build rather than stock vLLM. The causal audio tower starts from Voxtral Realtime 4B, and the weights are 8.17 GB. This is a different model from the Audio8 text-to-speech release in August.
