ANALYSIS Unitree Robotics is running two stories at once: a technical sprint toward a general-purpose humanoid brain and a market reckoning that has cut its valuation roughly in half, exposing the distance between China's humanoid robot ambitions and the commercial demand needed to justify them.
Why it matters
Unitree's trajectory since its August 19 STAR Market debut is the sharpest test yet of whether investor enthusiasm for humanoid robots can survive contact with actual revenue data. The company priced its IPO at 150.80 yuan per share and briefly touched 1,100 yuan intraday on its first trading day, lifting its market capitalisation to about 445 billion yuan, or roughly $66 billion4,2. By September 10, shares had fallen below 500 yuan, shrinking that figure to about 202 billion yuan. The online tranche was oversubscribed more than 8,200 times. ◆ That level of retail frenzy, followed by a collapse of more than 240 billion yuan in market value in roughly three weeks, is now prompting regulatory attention and raising questions about the entire sector's investment thesis.
The big picture
The core problem is commercial traction. Unitree's own filings show that 73.6% of its humanoid robot revenue comes from research and education buyers, not factories. Sales of more than 5,500 humanoid robots accounted for up to 868 million yuan of the company's 2025 revenue of 1.7 billion yuan. Net profit in 2025 was 278 million yuan. The company forecast first-half 2026 revenue of 1.052 billion to 1.128 billion yuan.
ANALYSIS Those numbers describe a profitable, growing hardware business, but not one that supports a $66 billion peak valuation built on the premise that humanoid robots are about to enter factories and homes at scale. The 73.6% research-and-education share is the critical detail: it means the robots are being bought to study, not to deploy.
Beijing's securities regulator has reportedly delivered informal "window guidance" to some investment banks and companies, tightening IPO scrutiny for humanoid robotics startups. Reuters reported on September 9 that this guidance was already in motion. ◆ The regulatory signal suggests authorities view the Unitree episode not as an isolated correction but as a symptom of broader speculative excess in the sector. More than 90 humanoid robot training centres are currently operating in China, and rival Agibot, which has not yet turned a profit, is pursuing its own listing at a high price-to-sales premium3.
Between the lines
The timing of Unitree's technical disclosure is striking. On September 10, the same day its shares slipped below 500 yuan, the company published the project page for UnifoLM-WLA-1.0, a roughly 6-billion-parameter humanoid foundation model trained on about 2,500 hours of real-robot data1. A single checkpoint coordinates 64 tasks spanning 10 whole-body and 54 tabletop manipulations, working with both two-finger grippers and multiple five-finger dexterous hands. The architecture layers an embodied-reasoning module called UnifoLM-ER-1, built on Qwen3-VL-4B and trained with more than 5 million embodied samples, atop a vision-language-action stack that uses optical flow, VQ-VAE compression, and residual vector quantization to handle the full action space.
Unitree claims UnifoLM-ER-1-4B leads open models on seven of 16 multimodal spatial and embodied-reasoning benchmarks, and beats proprietary GPT-6 Astra on specific spatial suites including Where2Place, BLINK spatial subtasks, CV-Bench, and EmbSpatial. Evaluation videos showed the model running on Unitree G1 hardware for chores such as folding towels, packing phones, making beds, and sorting shoes.
ANALYSIS The model release reads as an attempt to reframe the narrative from revenue shortfall to technological leadership. Yet the project page itself carries a caveat: code, models, and datasets still read "Coming soon," and weights and corpora are not yet downloadable.
The gap between Unitree and Agibot, as 36Kr framed it, maps onto two distinct bets: "hardware ontology" versus "AI brain". ◆ Unitree's simultaneous stock decline and model release crystallize that tension. The company is trying to prove it can do both, but the market is pricing the hardware reality, not the AI ambition.
What's next
About $900 million in IPO proceeds remains intact, giving Unitree a substantial runway to push UnifoLM-WLA-1.0 toward open release and to chase industrial customers beyond the research-and-education segment. Whether the window guidance from Beijing's regulator slows competing listings, including Agibot's, will shape the capital environment for the entire sector. The next concrete milestone is whether Unitree actually ships downloadable model weights, turning a project page into a platform others can build on.