A brokerage analyst in Shanghai needs two minutes of market commentary before the open. She types the script into a browser, picks a presenter from a grid of faces, chooses a studio background, and renders. The person on screen blinks, gestures at the right clause, and pronounces the ticker correctly.
Nobody involved has been to a studio. That is the product.
Xmov (魔珐科技), a Shanghai company founded in 2017 and usually described in its own materials as starting in 2018, has spent three years turning the hyper-real 3D avatar from a brand mascot into a piece of office equipment.
Key takeaways
- Xmov sells three products: 有言 for avatar-led video, 有光 for 24-hour livestreaming, and 有灵 for interactive agents, all built on a 3D avatar platform called 星云.
- The company says its library holds more than 3,000 hyper-real 3D avatars and that it works with over 1,000 enterprise clients across finance, government, retail, healthcare and education.
- Xmov claims a traditional hyper-real 3D avatar costs RMB 300,000–600,000 (≈ US$42,000–84,500 / HK$330,000–660,000) and takes months, while its pipeline renders in minutes.
- The company says it has raised about US$160 million in total, a figure that public funding databases do not fully corroborate.
- No revenue, margin, retention or customer-count data has been audited or published, so every commercial claim here is company-stated.
What the product is, stripped of the demo
Strip out the rendering and 有言 is a template engine. You supply a script, a presenter and a scene; the system supplies lip movement, gesture, camera moves and lighting. The output is a normal video file, not an interactive character.
有光 takes the same assets into a livestream, running unattended for hours with a preset script and question bank. 有灵 connects the avatar to a language model and a knowledge base, so the face can answer questions at a kiosk, a hospital desk or a government service hall. 星云 is the layer underneath, sold to developers who want to build their own.
The technical claim worth examining is the animation layer: most avatar tools animate only a face and a mouth. Xmov's pitch is full-body: gestures generated from the semantics of a script, so that a sentence about quarterly revenue produces a different movement than a sentence about a product defect. Whether that reads as natural or uncanny is an aesthetic judgment this article cannot settle from documentation.
The cost claim, and how to read it
The company's most repeated number is a cost comparison: a traditional hyper-real 3D avatar costs hundreds of thousands of yuan and takes months; its pipeline takes minutes.
The first half of that is broadly consistent with how the industry worked before generative tooling. A rigged, textured, performance-ready human model was a bespoke asset, built by a studio, priced like a bespoke asset. The second half is where the comparison gets slippery, because the two things being priced are not the same object. A bespoke avatar is an asset a company owns and can animate indefinitely. A rendered video is an output. Comparing the price of one output to the price of an asset flatters the output.
A fairer comparison is cost per finished minute over a year of use — which is also the comparison that favours Xmov when volume is high and disfavours it when a company needs one video. The company's own customer stories hint at this: the wins it publicises are all volume plays.
Who actually buys this
The case studies Xmov publishes share a shape. A company with many branches and one marketing team cannot produce enough video.
One eye-hospital chain is quoted saying it used 有言 to reach patients in bulk and cut content cost by 50%. Another customer, unnamed on the site, describes a marketing team that could not keep up with video demand across more than 700 stores. A consumer brand says it turned a printed poster into more than 300 product videos, so that non-hero products finally got a clip each. A provincial disease-control centre uses an avatar for public-health explainers.
That is a genuinely different market from the one digital humans were sold into five years ago. The buyer is not a brand paying for a virtual idol. The buyer is an operations manager with a content backlog and no camera crew. The avatar is not talent; it is throughput.
And throughput has a cost the case studies do not mention: sameness. When every branch of every chain uses the same small library of faces and studio backgrounds, the videos start to look like each other. The tool solves volume and creates a new problem — how to be recognisable when your competitor is using the identical template.
The funding numbers that disagree
Private-company reporting deserves a warning label, and Xmov is a good example.
Company materials, repeated across Chinese outlets, say the firm has raised about US$160 million from investors including ByteDance, SoftBank Vision Fund and Sequoia China. A funding database maintained by the research firm EqualOcean lists a US$110 million Series C in April 2022 led by SoftBank Vision Fund, a US$20 million B round in January 2023, and a total closer to RMB 1.05 billion (≈ US$148 million / HK$1.16 billion) across four rounds. The two accounts overlap but do not reconcile, and neither is audited.
None of this means the company is misrepresenting itself. Funding totals are aggregated differently — some count convertible notes, some only equity, some count announced rather than closed rounds. It means a reader should treat "US$160 million raised" as a public-relations figure, not a financial one.
The interesting threshold for avatar companies is not realism but whether the avatar can be handed to a line manager without a training course, because that is when the tool stops being a project and becomes infrastructure.
Honest limitations
This article is based on Xmov's own product pages and published customer stories, plus Chinese technology and business coverage; every commercial metric — clients, avatar library, savings, render times — is company-supplied and unaudited. We did not test the platform or view raw output. The funding figures conflict between company statements and a third-party database, and we could not obtain a definitive cap table. Claims about being cited in analyst reports are vendor statements we did not verify against the original documents. The cost comparison between bespoke avatars and rendered video compares different units of output and should not be read as a like-for-like saving.
What readers can do now
- Price avatar video per finished minute against your actual annual volume, not against the cost of a bespoke model.
- Ask who owns the avatar likeness and whether it can be reused if you switch vendors; template libraries are usually licensed, not sold.
- Pilot with your least sensitive content — internal training, not investor communication — before putting an avatar in front of customers.
- Treat funding totals from private Chinese AI companies as marketing until you see a filing, and prefer audited revenue to announced rounds.