The Hub is moving faster on vision speed, robot training, and messy speech. Liquid AI added a small vision drafter to its 3B model. LeRobot can now stream DROID video from Hugging Face storage with LanceDB. Basis released a large overlapping conversation set for multi-speaker research.
On the Hub
LiquidAI LFM2.5-VL-DSpark
Liquid AI posted about LFM2.5-VL-DSpark on September 24.1 The 279.5 million parameter vision drafter works with the existing 3B VLM, and the Hub repo has been live since September 18. Liquid reports decode speedups up to 3.13x on an M5 Max with MLX and 2.66x on an H100. The 3B target still checks every token and the license is LFM 1.0.
LeRobot native LanceDB
A Hugging Face community post on September 24 says LeRobot can now read Lance tables natively.2 Trainers can stream DROID video from Hugging Face Storage Buckets with a global shuffle instead of copying hundreds of gigabytes to local NVMe. On the full 369 GB DROID set, an eight-GPU H100 SmolVLA run of 10,000 steps finished in 1 hour 27 minutes from the bucket versus 2 hours from local disk, with the same final loss. Data-wait time fell from 37.4 percent to 1.7 percent.
Basis Conversations 1500
Basis AI published a Hub blog on September 23 about Basis Conversations 1500.3 The set is 1,502 hours of overlapping conversation with 2 to 4 speakers at once across 22 languages, with per-speaker 48 kHz FLAC tracks. The dataset card has been on the Hub since September 15 and about 100 hours carry human annotations. It is free for commercial and research use under the Basis data license and is built for messy multi-party speech.
