Unitree releases UnifoLM-WLA-1.0, a 6B humanoid foundation model

On October 3, 2026, a post on Reddit's LocalLLaMA community introduced UnifoLM-WLA-1.0, described as a new general-purpose humanoid foundation model from Unitree Robotics. The post is the only source in this account, and it links to a project page.

According to the post, the model has 6B parameters and was trained on roughly 2,500 hours of real robot data. A single model is said to handle 64 tasks: 10 whole-body tasks and 54 tabletop tasks. Hardware support covers parallel grippers and two different dexterous hands. On embodied benchmarks, the post says the model's spatial reasoning beats a lot of open-source models.

That is the full detail the report provides. It does not say when Unitree itself announced the model, how the training data was collected, which robot platform was used, or whether the model or its weights are available. The name UnifoLM-WLA-1.0 is given as-is, with no expansion of "WLA," and no benchmark scores or named comparison models are listed beyond the general claim about spatial reasoning. The post frames the model as a foundation model for humanoid robots, with one model presented as covering the listed tasks rather than separate task-specific policies.

No additional sources, official statements, or follow-up coverage appear in the provided material, so the release stands as a single community report rather than a confirmed timeline of events.