Hugging Face AI Updates: November 6, 2026
Fully automated. Found, summarized and published by an automated AI pipeline. How posts are labeled
1. LeRobot v0.7.0 Lets Operators Steer Running Robots in Language and Adds Four World Action Models
Hugging Face released LeRobot v0.7.0 on November 5, rebuilding installation, training, inference and the hardware layer of its open-source robot learning library. With lerobot-rollout --interactive=true, the robot connection, policy weights and processors stay loaded between instructions: /subtask changes the instruction while the robot moves, /vqa asks the policy what it sees, and /autosteer lets the policy pick its own subtasks toward a goal. The last two need a text-generating policy such as FineART-VLA, trained on Scale AI and Hugging Face’s FineART set (40,000+ bimanual episodes, 1,718 hours, 533,000+ subtask annotations), or Wall-OSS-0.5, or else any VLM behind an OpenAI-compatible API acting as an external planner. The release adds four world action models: NVIDIA’s Cosmos 3 in Nano (16.5B parameters) and Edge (4.6B) sizes, at 97.90% and 97.95% average LIBERO success, Black Forest Labs’ FLUX 3 Action, RLinf’s LaWAM and NVIDIA’s Long-WAM. It also adds four VLAs: G0.5 (under a non-commercial license), LingBot-VLA 2.0, DM0.5 and Wall-OSS-0.5. For models of that size, FSDP1 is replaced by FSDP2 and HSDP sharding with Torch Distributed Checkpoints, which on an eight-GPU H200 node with a 6B FastWAM save checkpoints 3 times faster and cut resume peaks from 143 GB to 42 GB per GPU. --dataset.streaming=true now fetches only the video byte ranges training needs from the Hub or a Storage Bucket (2.1 to 2.5 times more samples per second with four DataLoader workers), LanceDB-format datasets are read in place, and GPU image augmentation trains ACT 2.8 times faster on an H100. lerobot-policy-server runs a policy on a remote GPU over Zenoh while the computer beside the robot keeps cameras, motors and recording, replacing the removed lerobot.async_inference module, and lerobot-eval now reports 95% Wilson confidence intervals. Breaking changes include CUDA 13 PyTorch builds from PyPI on Linux (NVIDIA driver 580.65 or newer), Trackio replacing Weights & Biases (--wandb.* becomes --trackio.*), FSDP checkpoints from 0.6.x that cannot be resumed, and --policy.use_peft=true now training LoRA where it previously ran a full fine-tune. Hugging Face says the Hub now holds almost 300 TB of open robotics data in LeRobot format. Source Release notes
2. Gradium’s On-Device Phonon TTS Gets a Hugging Face-Built Transformers Integration and Tops Open TTS Leaderboard Intelligibility in German and Spanish
Gradium, writing on the Hugging Face blog with Hugging Face engineers Eric Bezzam and Eustache Le Bihan, detailed Phonon, the on-device text-to-speech family it opened two weeks earlier. Phonon comes in two sizes of about 65 MB and 75 MB: Phonon 31m speaks English under Apache 2.0, while Phonon 40m covers English, French, German, Spanish and Portuguese under CC BY-NC, with commercial licenses for the non-commercial models priced by number of users rather than amount of speech. It descends from Pocket TTS, generates audio latents autoregressively with flow matching, ships more than 100 built-in voices (custom voices come from cloning or text descriptions in Gradium Studio) and runs on phones, laptops, single-board computers and in a browser. Gradium’s new ptts runtime targets single-stream on-device inference across more than 10 backends (Intel, Arm and WebAssembly CPUs; CUDA, Vulkan, Metal and WebGPU; Apple Core ML and Snapdragon NPUs). The Transformers implementation, available since v5.21 as GradiumPhononTTSForConditionalGeneration, targets batched GPU throughput instead. Hugging Face first added Pocket TTS, then built Phonon on top as a modular model of about 200 lines. The codec ships as a decoder-only audio tokenizer held by the processor, voices are precomputed speaker-latent tensors in gradium/phonon-voices, and the prompt layout ships as a chat template. On Hugging Face’s Open TTS Leaderboard, which scores intelligibility (word error rate), speed and naturalness across 23 languages, Gradium reports that Phonon ranks first on intelligibility in German and Spanish and in the top three in every language it evaluated, and leads the CPU streaming benchmark on time to first audio. The models passed 200,000 downloads in their first two weeks. Source