What Is Spirit AI: Startup Changing How Machines Think and Work

Most robotics companies take a decade to ship something meaningful. Spirit AI did it in under two years. Founded in February 2024 in Hangzhou, China, the company is already running humanoid robots on live factory floors, topping global AI benchmarks, and raising hundreds of millions of dollars from some of the sharpest investors in the world. And people are paying attention.

So what exactly is Spirit AI, and why does it matter? Let’s get into it.

What Is Spirit AI? (And Why Everyone Is Talking About It)

Spirit AI, known in China as Qianxun Intelligence (千寻智能), is a robotics startup building what it calls a “Universal Brain” for robots. The idea is simple on the surface: instead of programming a robot to do one job, give it a general intelligence layer that lets it figure out new tasks on its own.

The reality is more complex. And more impressive.

The company was started by two people with very different but complementary backgrounds. CEO Han Fengtao spent over 12 years in industrial robotics. He co-founded and served as CTO of Rokae Robotics, where he led delivery of more than 20,000 robot units across 20-plus industries. That is not a lab background. That is real manufacturing, real problems, real deadlines. Co-founder Yang Gao comes from the research side. He studied at UC Berkeley under Pieter Abbeel, one of the most respected names in embodied intelligence, and his work has appeared at top conferences like NeurIPS and ICRA.

The rest of the team? Mostly under 30. Drawn from UC Berkeley, Tsinghua University, Peking University, Carnegie Mellon, ByteDance, Xiaomi, and DJI. It reads like a who’s-who of Chinese and global AI talent squeezed into one startup. In 2025, The Information named Spirit AI the No. 2 most promising startup in Asia on its list of 50 top global startups, selected from tens of thousands worldwide.

How Spirit AI Robots Work in the Real World

Here is where it gets genuinely interesting. Most industrial robots are rigid. They follow scripts. If a component shifts by two centimeters, a traditional robot fails. Spirit AI’s robots do not work that way.

The core of their technology is a Vision-Language-Action model, commonly called a VLA model. Think of it as three things fused together: the robot sees its environment, understands instructions in natural language, and then takes physical action. Not a pre-written action. A reasoned one.

Their humanoid robot, the Moz1 (nicknamed “Xiao Mo”), was officially released in June 2025. It carries 26 degrees of freedom across its entire body. That level of articulation means it can handle the kind of precise, continuous manipulation tasks that have historically defeated machines. It runs two proprietary AI systems: ViLa, a visual language model that lets it perceive and interpret its surroundings, and CoPa, a component constraint model that helps it understand how physical objects relate to each other.

The numbers back it up. In real-world deployment, the Moz robot achieved a task success rate consistently above 99%. That is not a demo stat. That is from an active production line.

Spirit AI’s “Universal Brain” Explained Simply

Let’s be honest. Most people’s eyes glaze over when they hear terms like “foundation model” or “zero-shot generalization.” So here is the plain version.

Today’s robots are specialists. A robot that handles car parts cannot fold laundry. A robot trained for one factory cannot walk into a different one and figure things out. They are expensive, single-purpose machines. And that limits where and how fast they can be deployed.

Spirit AI’s answer is the Spirit v1.5 VLA model. It is designed to work like an operating system for robots. One brain, many bodies, many tasks. In January 2026, Spirit v1.5 ranked first globally on RoboChallenge, a real-world benchmark widely regarded as the most rigorous test of embodied AI. It scored 66.09 overall with a task success rate of 50.33%. It was the only model in the world to break the 50% threshold.

And here is the kicker. That score beat pi0.5, the top model from US-based Physical Intelligence, a company backed by Silicon Valley’s heaviest hitters.

Spirit AI then did something unexpected. They open-sourced the model. Released the weights, the evaluation code, everything. A direct invitation to the global research community to verify the results and build on top of them.

Spirit AI vs. Other Humanoid Robot Companies

The humanoid robot field is noisy right now. Tesla Optimus gets the headlines. Boston Dynamics has the legacy. Figure AI has the $39 billion valuation. Unitree Robotics has the affordable price tag. So where does Spirit AI actually stand?

The honest answer is: differently positioned, and in some ways further along.

Boston Dynamics was founded in 1992. It has 32 years of hardware engineering behind it and the full weight of Hyundai’s manufacturing ecosystem. Figure AI is valued at more than any pure robotics startup in history. These are serious competitors.

But Spirit AI, founded in 2024, has already deployed robots on live production lines inside CATL, the world’s largest electric vehicle battery maker. Not a pilot program. Not a press release. Active deployment. That is a harder claim for most of its Western counterparts to match right now.

The strategic partner lineup also tells a story. Spirit AI counts CATL, JD.com, Huawei, Xiaomi, and TCL among its industrial partners. That network gives it something money cannot easily buy: access to real deployment environments and operational data at scale.

The focus is also sharper. Where some companies are trying to build robot bodies that can dance or do backflips, Spirit AI is focused on robots that can work. Messy, repetitive, high-precision industrial tasks that actually need solving.

Spirit AI Funding: How It Raised $280 Million

In February 2026, Spirit AI raised approximately $280 million USD across two back-to-back funding rounds, bringing total funding to nearly 2 billion Chinese yuan. Fast. Very fast for a company that did not exist two years prior.

The investor list is not a random collection of venture funds. Yunfeng Capital, Sequoia China (HongShan), Chaos Investment, Prosperity7 Ventures, Synstellation Capital, and TCL Capital all participated. These are institutions that see a lot of deals. They do not write checks this size for companies they are not sure about.

The capital is earmarked for three things: scaling the AI model deployment, expanding the data pipeline from its current 200,000 hours of interaction data to over 1 million hours by end of 2026, and accelerating commercial robot rollouts in manufacturing and logistics.

One detail that rarely gets highlighted but matters a lot: Spirit AI built its own proprietary wearable data collection devices, now in their fifth generation. By having humans wear sensors and perform natural tasks, they collect training data at roughly one-tenth the cost of traditional teleoperation methods. That is not a minor efficiency gain. That is a structural cost advantage that compounds as they scale. Spirit AI projected it would exceed 100 million yuan in revenue in 2026, with robot shipments reaching several hundred units.

Key Features of Spirit AI’s Technology

So what actually makes up the technology stack? Here is what the sources confirm.

Spirit v1.5 VLA Model. The AI core. Ranked first globally on RoboChallenge in January 2026. Open-sourced the same month. Built for zero-shot generalization, meaning the robot can handle tasks it has never seen before without additional training. That is the hard part. Most models fail here.

Moz1 Humanoid Robot. Released June 2025. Described as China’s first high-performance full-body torque-controlled humanoid robot. Twenty-six degrees of freedom, self-developed joint motors with leading power density. Designed for industrial tasks requiring continuous precision.

ViLa and CoPa AI Models. ViLa handles visual and language understanding. CoPa manages component-level physical constraints. Together they power the robot’s ability to adapt in real time without stopping for recalibration.

Wearable Data Collection System. Fifth-generation proprietary hardware. Reduces data acquisition costs to one-tenth of traditional methods. The competitive advantage here is not just cost. It is speed of data accumulation.

Full-Stack Build. Spirit AI builds both the AI software and the robot hardware in-house. That is rarer than it sounds. Most companies outsource one or the other. Full-stack control means faster iteration and tighter integration between brain and body.

What Spirit AI Means for the Future of Work

On December 17, 2025, something happened in Luoyang, China that did not get nearly enough coverage in Western media.

Spirit AI’s Xiao Mo robots went live on CATL’s Zhongzhou battery production line. Not as a test. As workers. Their assigned task: plugging high-voltage connectors into battery packs during end-of-line testing. It sounds simple. It is not.

The packs carry hundreds of volts. The wire harnesses are flexible, meaning they shift unpredictably. For human workers, the task carried real electric shock risks. For traditional robots, the unpredictability of the flexible cables made it nearly impossible to automate reliably. It had been a persistent bottleneck.

Xiao Mo handled it. A task success rate above 99%, a threefold increase in daily workload compared to human workers, and exceptional consistency across multiple battery models. CATL put it plainly: “Moz has become an indispensable member of our production line.”

And that sentence carries more weight than any benchmark number. An indispensable member. Not a machine running alongside the team. A member of it.

The bigger picture is this. Spirit AI’s target is to empower 10% of the world to own a robot within 10 years. That is an ambitious number. But companies that set audacious targets and then show up on CATL’s factory floor ahead of schedule deserve to be taken seriously.

The future of work does not look like science fiction anymore. It looks like Hangzhou, 2024, two founders, one big idea, and a robot named Xiao Mo doing a job humans were too much at risk to do themselves.

Spirit AI Business Model