Unitree open-sources a 6B humanoid brain that beats GPT-6 Astra
Robot makers have spent two years shipping better bodies than brains. This week one of them open-sourced the missing half.
Unitree released UnifoLM-WLA-1.0, a 6B-parameter humanoid foundation model trained on about 2,500 hours of real-robot data that runs 64 tasks from a single checkpoint — 54 desktop, 10 whole-body — across two-finger grippers and several five-fingered dexterous hands, and it published the weights, code and datasets instead of a demo reel. The embodied-reasoning base is UnifoLM-ER-1-4B, built on Alibaba's Qwen3-VL-4B and trained on more than 5 million embodied-reasoning samples. According to the writeup in Chinese outlet Zhidi, that model leads 7 of 16 multimodal perception and understanding benchmarks among the open models tested, and on spatial reasoning it edges past GPT-6 Astra: 82.0 against 69.0 on Where2Place, 88.9 against 83.3 on EmbSpatial.
The design bet is that predicting what changes beats predicting the whole next frame. Optical flow isolates the regions an action actually alters — the arm, the object, the contact patch — and those regions are compressed into a fixed string of discrete tokens, which is what Unitree calls interaction-centric world modeling. The action space is split three ways (end-effector pose, end-effector joints, lower-body joints), each encoded separately so different parts of the body can be learned on one timeline, and a final expert module decodes the tokens back into continuous motion the hardware can execute. The published task list is deliberately mundane: folding towels, loading dishes, charging batteries, making beds, taking out the trash, loading a washing machine.
That homeliness is the argument. A model that only works in one scene is a demo; a checkpoint that spans 64 tasks is a bet on a general-purpose embodied stack, and Unitree is positioning it as the Android of robot brains — the layer everyone else adapts. The 2,500 hours is the part rivals can't copy quickly: that is one robot recording continuously for more than 100 days, and unlike text, real-robot data can only be produced by running real robots. It also answers the criticism we made in August, when Unitree's valuation was halved on the view that its brain had not caught up with its body — The Take — Unitree's rout isn't a bubble. It's the brain lagging the body. Open weights don't weaken that moat, they push the competition onto data and hardware, where Unitree sells physical machines.
More than 100 AI evaluators and researchers signed a public letter on Friday spelling out what "independent evaluation" has to mean in practice — funding that survives unflattering findings, direct access to boards, and explicit protection from retaliation. The letter, organized by the AI Evaluator Forum and shared exclusively with CNBC, is a response to lab pledges: Anthropic CEO Dario Amodei floated giving outside evaluators "employee-like access" over the weekend, and Sam Altman, Elon Musk and Microsoft's Satya Nadella have backed the idea — Altman matches Amodei's evaluator pledge, Musk says 'Dario is right'. Signatories include Geoffrey Hinton and researchers from Johns Hopkins, Stanford and the nonprofit evaluator METR.
The five conditions are specific: evaluators must not be owned or governed by a lab, must not take contingent payment, must be multiple organizations across different risk areas, must have publication rights on a limited redaction clock, and must get access equal to highly privileged internal staff — including candid one-on-ones. Forum chair Conrad Stosz cited the unreleased OpenAI model involved in the Hugging Face attack as the reason internal systems need outside eyes. The same standard, AEF-1, already has EU AI Office endorsement, so the letter is less a plea than a compliance template.
Cohere CEO Aidan Gomez says the U.S. story that China's AI progress is just distillation fails on the benchmark sheet — "you can't copy or distill to better." On CNBC's The Tech Download, Gomez, a co-author of the 2017 paper behind most of today's AI services, called Chinese models "world class" and said the U.S. lead is "evaporating very quickly," while conceding distillation happens to some extent. Former White House AI policy adviser Sriram Krishnan made the same point on air: ChatGPT and Claude "came out of distilling human content."
The counter-case still stands — Anthropic alleges Alibaba, Moonshot and DeepSeek ran illicit distillation campaigns, CISA calls it the core of China's strategy, and Beijing's commerce ministry calls the allegations "groundless and legally unsound" — US agencies call distillation the core of China's AI strategy. What's new is that the doubters now include people selling Western models, and their explanation for China's progress is duller and harder to legislate against: chip restrictions forced leaner architectures.
What to watch: whether anyone reproduces Unitree's spatial scores on hardware and scenes the company never recorded.
If a robot brain ships as open weights, does the data pipeline behind it become the real product? Tell us in the comments.
Sources: Unitree UnifoLM-WLA-1.0 (GitHub)
