Unitree Robotics published UnifoLM-WLA-1.0 on September 10, 2026, a humanoid robot foundation model with about 6 billion parameters that the company says was trained on roughly 2,500 hours of real-robot data and runs 64 manipulation tasks from a single checkpoint. The release was widely described as open source, but at announcement the project page marked the code, model weights and datasets as "coming soon."

Why it matters: Unitree is best known for low-priced humanoid hardware. UnifoLM-WLA-1.0 is its most detailed attempt to ship the model layer as well, and the gap between the announcement and what can actually be downloaded is the part to watch.

What is UnifoLM-WLA-1.0?

According to the project page, UnifoLM-WLA-1.0 is a 6B-parameter vision-language-action (VLA) model. It sits on a multimodal backbone called UnifoLM-ER-Flow, which builds on UnifoLM-ER-1, a 4B vision-language model, and adds an MMDiT action expert that produces three kinds of output: end-effector poses, end-effector joints and lower-body joints. Unitree lists more than 5 million embodied-reasoning samples in its training mix alongside the robot data. The GitHub README credits the open starVLA and Qwen-Image projects as the base it builds on.

The model supports two-finger grippers and several five-finger dexterous hands. The Dex1 datasets Unitree has posted alongside it are labelled for its G1 humanoid.

Key Facts

  • Announced September 10, 2026; about 6 billion parameters
  • Trained on about 2,500 hours of real-robot data plus 5M+ embodied-reasoning samples (company-reported)
  • 64 tasks from one checkpoint: 10 whole-body, 54 tabletop
  • Claims top open-model results on 7 of 16 embodied-reasoning benchmarks (company-reported)
  • As of September 23, 2026: 4B backbone weights and action-expert code released; full UnifoLM-WLA-Base checkpoint not yet released

What can it do, according to Unitree?

Unitree reports 64 tasks from one model: 10 whole-body manipulation tasks and 54 tabletop manipulation tasks. On embodied reasoning, the company says the model leads open models on 7 of 16 multimodal perception and embodied-reasoning benchmarks and is comparable to leading proprietary models overall. All of these figures are company-reported, taken from Unitree's own project page. Let's Data Science, reviewing the launch, called it "best understood as a detailed project announcement rather than a reproducible open release."

Is UnifoLM-WLA-1.0 actually open source?

Partly, and only after launch day. The project is licensed Apache 2.0. The release log on GitHub shows three steps so far:

As of September 23, 2026, the repository's own checklist still shows the UnifoLM-WLA-Base checkpoint, fine-tuning code and LoRA fine-tuning code as not yet released. Three datasets are marked available: the UniBot-V1 Challenge Dataset, the UnifoLM-WBT-Dataset and the UnifoLM-Dex1-Dataset. In other words, developers can now train their own action expert on Unitree's backbone, but cannot yet run the 6B model that produced the 64-task results.

For background on how VLA models split perception, language and motor control, see our VLA explainer.

Update, September 23, 2026: this article was updated to reflect the September 11 and September 20 releases.

Frequently Asked

What is Unitree UnifoLM-WLA-1.0?

UnifoLM-WLA-1.0 is a humanoid robot foundation model with about 6 billion parameters that Unitree Robotics published on September 10, 2026. Unitree says it was trained on about 2,500 hours of real-robot data and runs 64 manipulation tasks from a single model.

Are the UnifoLM-WLA-1.0 weights available?

Only partly. At announcement, code, weights and data were marked "coming soon." Unitree posted the UnifoLM-ER-1 and UnifoLM-ER-Flow backbone weights on September 11 and action-expert training code on September 20, but as of September 23, 2026 the full UnifoLM-WLA-Base checkpoint was still listed as unreleased.

What tasks can UnifoLM-WLA-1.0 perform?

Unitree reports 64 tasks: 10 whole-body manipulation tasks and 54 tabletop manipulation tasks, using two-finger grippers and several five-finger dexterous hands. These results are company-reported.

How does UnifoLM-WLA-1.0 compare with other robot models?

Unitree says the model posts the best results among open models on 7 of 16 multimodal perception and embodied-reasoning benchmarks and is comparable to leading proprietary models overall. Those comparisons come from Unitree's own project page.

Sources & Further Reading