Meta built a humanoid robot prototype with Chinese startup Fourier and is now working on a second version, according to a report from The Information that Reuters carried on October 31, 2026. Reuters said it could not immediately verify the story. The same coverage says the prototype is codenamed Max, weighs about 106 pounds, and was assembled in roughly three months. Meta's stated position, per the reporting, is that it has no plans to sell these units.
That is the whole confirmed-by-someone picture, and it is thin: one primary report, a Meta line about not selling the machines, and no public specs, demo or paper. This post separates what is claimed from what is verified, adds the history that makes the story make sense, and says how to judge any later announcement.
TL;DR: claimed vs verified
| Question | Answer |
|---|---|
| Who reported it? | The Information, carried by Reuters on October 31, 2026 |
| Did Reuters verify it? | No. Reuters said it could not immediately verify the report |
| What is the prototype called? | Max, a codename, per secondary coverage of the report |
| How heavy is it? | About 106 pounds (roughly 48 kg), same source |
| Who built it? | Meta with Fourier, over about three months, per the report |
| Will Meta sell it? | Meta says it has no plans to sell these units; no consumer timeline |
| Is there a public demo, spec sheet or paper? | None found from Meta or Fourier |
| What AI runs it? | Not stated; one outlet says Meta sees humanoids as an embodiment of its Muse agent, which is not a confirmed technical detail |
What exactly is being reported
According to Reuters' write-up of The Information's story, Meta has built a prototype humanoid with Fourier and is developing a second iteration. Outlets repeating the report add three details: the prototype took about three months to build, it weighs about 106 pounds, and it is internally called Max. One summary says Meta views humanoids as the physical form of its Muse agent and imagines household tasks as the goal. We treat that last line as the outlet's framing, because no source we could open quotes Meta saying it.
Two things are conspicuously missing. There is no description of how the robot walks or manipulates objects, and there is no named model controlling it. A humanoid frame is the easy part to build with a partner; the hard part is the AI that lets it do useful work, and that is exactly what the report does not describe. For a sense of how hard that is to measure, see our look at an AA Robotics benchmark of frontier models driving robot arms and humanoids.
Why "will not be for sale" is the key line
The most useful sentence in the story is the one that limits it. Meta's statement that it has no plans to sell these units fits what was reported in February 2025, when Bloomberg-sourced coverage said Meta was setting up a humanoid effort inside Reality Labs. The Decoder summarized that memo at the time, and TechSpot's account noted the stated focus on household chores. Search results from that period also describe a plan to supply AI, sensors and software to companies that build the hardware, rather than ship a Meta-branded robot first.
So the likeliest reading is a research and learning platform, not a product. A prototype built with an experienced hardware partner lets Meta's teams collect data and test models on a real body while the company decides whether to license software, partner on hardware or build its own. That is an inference from the history, not something Meta has said about Max.
Who is Fourier, and why does a Chinese partner raise eyebrows?
Fourier, formerly Fourier Intelligence, was founded in 2015 by Alex Gu, according to its Wikipedia entry. It began in rehabilitation robotics and later built the GR-1 humanoid, a program that started in 2019. It is one of several Chinese makers, alongside Unitree and others, that have pushed down the cost of humanoid hardware.
The reporting headline stresses that the partner is Chinese. That matters for a US company because of export controls, data-handling rules and political scrutiny of China-linked hardware. Nothing in the sources we could read says how the work was structured, where the prototype was built, or what data it touches. Those are open questions, and we do not guess at answers here.
How this fits Meta's AI push
Meta has been tying its AI agent products to devices for months. Our coverage of Meta's Muse family and the proactive-agent race and of how Meta's Muse apps treat private data shows an agent strategy that wants more access to your life. A humanoid is the extreme version: a camera and microphone array that moves through a home.
That is why the privacy and safety questions come before the capability ones. A home robot that runs a company's agent would need to say where video goes, what is stored, and who can see it. Meta has published none of that for Max, because Max is not a product. The company's relationship with Apple on local Muse inference, covered in our Apple and Meta Muse Spark partnership post, hints at the on-device direction, but nobody has tied that to robotics.
What the rest of the field looks like
Humanoid and robot-foundation-model news is arriving weekly. A useful comparison for how a small team can ship an embodied model through an API is Dream Labs' Dream 1 world-action model, which publishes a price and an interface. Compare that bar to Max: no price, no interface, no demo, no benchmark. That does not mean Max is unimpressive; it means there is nothing to judge yet.
When a company does publish, these are the things worth checking:
- A video of unedited runs, not a teleoperated stage demo. Ask whether a human was driving.
- Task success rates over many trials, not a single successful pick-and-place.
- Which model controls the body, and whether it is trained on the company's own data.
- Safety behavior: what happens when the robot loses balance near a person.
- Cost and supply: who manufactures the actuators, and what each unit costs.
Timeline: how Meta got here
| Date | Event | Source |
|---|---|---|
| February 2025 | Memo-based reports say Meta is forming a humanoid robotics effort inside Reality Labs, aimed at household tasks | Bloomberg-sourced coverage, via The Decoder and TechSpot |
| 2025 | Coverage describes a plan to supply AI, sensors and software to hardware makers rather than sell a Meta robot first | Outlet summaries of the same reporting |
| October 31, 2026 | The Information reports a Fourier-built prototype, Max, and a second iteration in progress | Reuters, citing The Information |
The pattern is consistent: Meta has talked about humanoids as a platform bet for years, while saying little about hardware it would own. The Max report is the first time a specific body, partner and weight have been attached to the effort, and all of it rests on one outlet's sourcing.
What a three-month build does and does not tell you
Building a 106-pound humanoid in three months is plausible only if most of the mechanical work comes from a partner with an existing platform, which is how the report frames it. It says little about the intelligence layer. A robot arm that reaches, a leg that balances and a model that understands a request like "tidy the living room" are three separate problems, and the last is the one Meta's AI labs would own.
For readers choosing where to learn this field, the useful takeaway is the split between body and brain. Hardware is increasingly a purchasable commodity, as Chinese makers have shown, while the controlling model is where the differentiation and the data collection live. That is why the "not for sale" line reads less like a retreat and more like a statement about where Meta intends to compete.
What developers and readers should take from this
- Treat the numbers as the reporter's, not Meta's. The 106-pound and three-month figures appear in one report and its echoes. If Meta posts a spec, update accordingly.
- Do not read "prototype" as "product". The company said it does not plan to sell these machines.
- Watch for the second iteration. Any official video or paper about version two would be the first real data point.
- Watch Reality Labs and Meta AI research channels for a paper or open model tied to embodied control. Meta has a history of releasing research artifacts, and a robotics release would be easy to verify.
Open questions nobody has answered
- Is Max controlled by a Meta-trained model, a Fourier controller, or both?
- Is the second version a refinement of the same frame or a new design?
- Does the project have a publication or open-source plan?
- How does Meta handle the geopolitical side of working with a Chinese hardware partner?
Until one of these is answered by Meta or Fourier directly, the accurate summary is simple: Meta is reported to have a humanoid prototype, Reuters could not confirm it, and Meta says it will not sell it. We will add an update section here if either company publishes specs, a demo or a statement.
Details reflect reporting available on October 31, 2026 and may change.
Related reading
- AA Robotics v0.1: frontier models on robot arms and humanoids, claimed vs verified
- Dream Labs Dream 1 robot world-action model API
- Claude Duet vs OpenAI Dots vs Meta Muse vs Grok bot: proactive agents
- AI agent privacy promises: Meta Muse, OpenAI Dots and the Verge analysis
- Apple and Meta Muse Spark Open on Mac Studio
- Dynamic Creatures Hugo companion robot, claimed vs verified