A Body for Muse: Meta's 'Max' Humanoid Was Built With China's Fourier in Three Months

Meta told the world in February 2025 that it was forming a humanoid robotics group, and then said almost nothing for twenty months. On Saturday 31 October 2026, The Information’s Jyoti Mann reported, citing people familiar with the project, what that group has actually been doing: it built a working humanoid prototype with Fourier, a Shanghai robotics company, in roughly three months, showed it to staff in August, and is already on the second version. Reuters picked the story up the same day; The Next Web, which summarised it on 1 November, noted it had not independently verified the details. Meta’s own on-record comment is narrow: the Fourier work is “focused on internal research and experimentation”, and it has no plans to sell these units.

That caveat matters, because almost everything below rests on one outlet’s sourcing. But the shape of what was reported is consistent, and it tells you more about Meta’s robotics strategy than anything the company has said itself.

What was reported

The prototype is known internally as Max. It stands about 1.5 metres tall and weighs about 48 kg, runs on NVIDIA’s Jetson Thor robot computer, has five-fingered hands said to be capable of picking up small bottles and toys, chest-mounted speakers, and a screen for a face that can display expressions. It was built as a custom machine on Fourier hardware, drawing on the three generations of humanoids Fourier has shipped, with the head, display and cameras redesigned to Meta’s specification.

The human side of the story is Li-Chen Miller, Meta’s vice president of product for robotics, who reportedly spent almost two months working at Fourier’s factory in Shanghai. The collaboration ended in an internal demonstration at a Meta robotics summit in late August. Work on Max 2 started in September: a redesigned torso, improved hands, a projector, and additional sensors meant to help the robot handle stairs. Meta is also said to be hiring robotics staff in Shanghai and Shenzhen. No launch date for any humanoid product has been set.

The spec sheet is about conversation, not chores

The most revealing detail is the list of requirements Meta reportedly gave Fourier. Max had to recognise a voice from five metres away, interpret a wave or gesture from three metres, maintain eye contact, and signal clearly whether it is listening, thinking, speaking or uncertain.

Read that against what other humanoid makers advertise. Figure, Tesla and the Chinese industrial players sell manipulation: picking, placing, sorting, assembling. Fourier’s own GR-3 is marketed as a “care-bot” for companionship and eldercare. Meta’s list is a social-interaction spec. The hands are modest, the face is a screen, and the four interaction states are exactly the status cues a chat interface shows you while an agent works. Max, at least in its first version, is an interface device: a smart speaker with legs and a face.

That reading matches the reported framing inside Meta. CTO Andrew Bosworth has described humanoids internally as the physical embodiment of Muse, the personal AI agent Meta launched in the United States on 8 September 2026. Muse runs on the Muse Spark model, executes tasks inside a dedicated “Secure VM” with its own browser, and is “coming soon to AI glasses”, in Meta’s words. On 27 October, NVIDIA agreed to put 3 billion USD into Meta’s share offering as part of an expanded partnership to train and serve the Muse models. A week later we learn the agent is being fitted with a body. The sequence is not a coincidence.

For developers there is a concrete takeaway in the “listening, thinking, speaking, uncertain” requirement. It is the same problem as showing a spinner or a tool-call trace in a chat UI, but with no screen to hide behind. Whoever ends up writing software for robots like this will spend a surprising share of their time on state signalling and turn-taking, not on grasping. The compute is also instructive: Jetson Thor is NVIDIA’s Blackwell-based robot module with 128 GB of memory and a headline figure of 2,070 FP4 TFLOPS (sparse), sold as a 3,499 USD developer kit since 2025. It is enough to run perception and a mid-sized language or vision-language-action model on the robot itself, and it is the same part NVIDIA’s partners used for the reference humanoid announced at GTC Berlin on 21 October. Max is built from off-the-shelf 2026 robotics parts.

Why the hardware came from Shanghai

Fourier is not a random choice. The company was founded in Shanghai in 2015 by Alex Gu, started in rehabilitation robotics (its first product was the X1 walking exoskeleton in 2017), began humanoid work in 2019, and unveiled the GR-1 at the World Artificial Intelligence Conference in July 2023. The GR-2 followed in September 2024 (1.75 m, 63 kg, 53 degrees of freedom), and the GR-3 in August 2025 (165 cm, 71 kg, 55 degrees of freedom, a 3 kg single-arm payload). Its backers include IDG Capital and Saudi Aramco. Three shipping generations in three years is precisely the kind of integration experience a software company lacks, and the reporting says Max leans on it.

The broader point is where the humanoid supply chain now lives. Barclays has estimated that China accounted for 85 per cent of the world’s humanoid robot installations in 2025, and the Shanghai firm AgiBot claimed the top spot in global shipments in the first half of 2026. Actuators, dexterous hands, battery packs and the engineers who have already debugged a bipedal chassis are concentrated in Shanghai and Shenzhen. A US company that wants a working body in a quarter, rather than a programme, goes there.

That is also consistent with how Meta described its intentions from the start. When Bloomberg first reported the group in February 2025, led by former Cruise chief executive Marc Whitten inside Reality Labs, the plan was explicitly not to build a Meta-branded robot, at least initially, but to build a hardware and software foundation that other manufacturers could adopt, in the way Android did for phones. Meta was said to be in talks with Unitree and Figure about prototypes. In November 2025 Li-Chen Miller, who had run the smart-glasses portfolio, became the first product manager of the robotics team, which by then had nearly 40 open roles and an internal goal of a household humanoid nicknamed “Metabot”. In May 2026 Meta bought Assured Robot Intelligence, a San Diego startup founded by the researchers Lerrel Pinto and Xiaolong Wang; its team joined Meta Superintelligence Labs, which also houses the Meta Robotics Studio that leads the humanoid effort. Max is the first physical evidence that these pieces are being assembled into one thing.

What three months does and does not prove

The headline, that Meta built a humanoid in three months, deserves some deflation. What was built in three months is a customised Fourier robot with a new head. The achievement is mostly Fourier’s integration speed and Meta’s product requirements, not a new machine. “Meta built” is generous.

More importantly, nothing in the reporting speaks to autonomy. There are no task success rates, no statement of whether the August demo was teleoperated or scripted, and no mention of which model, if any, was driving the body. Compare that with what a serious robotics announcement looks like now: AgiBot’s GO-3 came with benchmark placements and learning curves, and the Stanford HomeBody work reported success rates on unseen kitchens. Meta’s May acquisition brought in people who build whole-body control models; whether their work is running on Max is unknown. The fact that stairs are on the Max 2 wish list tells you Max cannot climb them.

And it is a single-source story. Reuters and others relayed it with attribution to The Information. Meta confirmed the existence of the Fourier work and its research-only status, and nothing else. Treat the specifications and the timeline as reported, not established.

Why it matters

Even discounted, the report settles two questions. First, the largest consumer software company on the planet has decided that its agent needs a body, and that the body should live in homes. That is a much bigger bet than selling robots to factories, and it comes from the company whose Reality Labs division has spent a decade and tens of billions learning how hard consumer hardware is.

Second, the brains-versus-bodies split is becoming explicit. US companies are building the models; Chinese companies are building the machines; and the first thing a US giant does when it needs a working humanoid is fly its product lead to Shanghai. If a Meta robot ever ships, the question of whether a camera- and microphone-laden home device can be built on a Chinese chassis will attract the same political scrutiny that other Chinese-made connected hardware has drawn in Washington. Meta’s “no plans to sell these units” is, among other things, a way to avoid that question for now.

For the IT industry, the platform play is the thing to watch. If Meta follows its own Android analogy, the product will be a robot operating layer and the Muse agent on top, licensed to manufacturers like Fourier, with the hardware as a reference design. That would make the robot’s API surface the agent’s API surface, and it would put the social-interaction layer, the part Max was specified around, at the centre of robotics software rather than at the edge.

Open questions

Ship or license? Meta’s reported plan since 2025 has been to be the platform, not the manufacturer. Max 2 and hiring in Shenzhen could mean either.
Whose model drives it? The Assured Robot Intelligence team was brought in for whole-body control; nothing reported says their models run on Max.
Which organisation owns it? Reporting places the robotics group in Reality Labs and the Robotics Studio in Meta Superintelligence Labs. Hardware and models under different roofs is a familiar Meta pattern, and not a happy one.
Privacy. A humanoid that watches for a wave from three metres away is a mobile camera in the home. Muse already drew trust questions as software; embodied, those questions multiply.
Policy. A Chinese-built body for an American agent is viable for research. Whether it is viable as a product depends on decisions that are not Meta’s to make.