Unitree open-sources 6B humanoid model UnifoLM-WLA-1.0, claims fine-tuning on one 24GB GPU
lmoroney · x · 2026-10-12
Unitree released UnifoLM-WLA-1.0, a 6B-parameter humanoid model open-sourced under Apache 2.0 on Hugging Face (12.5 GB checkpoint), said to drive its G1 humanoid through 64 tasks (54 tabletop, 10 whole-body).
Key points:
- The default fine-tuning recipe freezes the vision-language backbone and trains only the action head, which Unitree says fits on a single 24 GB GPU
- The release rolled out in three steps: two embodied-reasoning models (UnifoLM-ER-1, UnifoLM-ER-Flow) on September 11, then the weights and fine-tuning code
- Unitree labels the demo video's whole-body clips as autonomous and not sped up
Notably, Unitree has not published success rates or benchmark numbers — a demo reel isn't a benchmark. Still, open weights from a company that has shipped thousands of humanoids are valuable to researchers and robot startups.
Related event: Unitree Open-Sources 6B Humanoid Robot Model UnifoLM-WLA-1.0(2 posts)→
More from Embodied
- Vernata: ETH Zurich's multi-teacher distillation for self-supervised LiDAR representations, IROS 2026 — rsasaki0109 · 2026-10-12
- Zero real-world data: Success-Guided Sampling unlocks zero-shot dexterous robot manipulation via sim2real RL — Scobleizer · 2026-10-12
- Dev asks Grok to build itself a Vision Pro app, creates conversational room bots — Baconbrix · 2026-10-12
- Local AI RPG stack combines 3 DGX Sparks, GLM 5.3 Flash and ComfyUI — HankYeomans · 2026-10-12
- AI Agent Does CAD: KiCad MCP Reads Mounting Holes for a Robot Build — burhop · 2026-10-12
- Siemens Brings Nvidia Omniverse into Digital Twin Composer to Pave the Way for Physical AI — RevLebaredian · 2026-10-12