Xmax AI Releases X2.0: Videos Are No Longer Just Watched, but Played

Early this morning, Xmax AI dropped its new generation general-purpose real-time interactive video model, Xmax X2.0, on X (Twitter), along with open access to its commercial API and industry collaboration channels. It has been only 5 months since they released X1 in February.

If you hadn’t heard of this company last year, that’s fine — after all, in China, they are currently the only team that has managed to commercialize and scale real-time interactive video technology successfully.

The Bottom Line: What’s New in X2.0

In a nutshell: X1 made people realize for the first time that “videos can be played”; X2.0 turns that into infrastructure.

Two major upgrades:

• Image quality improved from 480p to 960p, 24fps — sufficient for consumer-grade use cases like live streaming, social, and e-commerce.
• Latency reduced to the millisecond level, virtually zero lag. That’s the real game-changer.

Calling this just an “experience optimization” would be an understatement. Users go from “entering a prompt and waiting for AI to render” to “directly controlling the video in real time.” These are fundamentally different products. The former is a tool, the latter is an interface — the difference between “using Photoshop” and “looking in a mirror.” Once waiting disappears, the entire mental model of interaction changes.

Under the hood is an end-to-end streaming re-rendering architecture, where the model performs autoregressive generation frame by frame; each frame’s output doesn’t depend on rendering the entire preceding sequence. In other words, it doesn’t plan the whole video in advance — it renders and responds in sync, frame by frame.

Unified Interaction Architecture: Beyond Just Prompts

For the past two years, the AI video race has revolved around a quality arms race. From Sora to Seedance, Kling, and Veo, by mid-2026, the visual quality gap among top-tier models will be almost indistinguishable to the naked eye.

Xmax took a different route: instead of competing on quality, it competes on interactivity.

X2.0 introduces a unified multimodal interaction architecture, enabling one model to simultaneously understand:

• Touchscreen dragging
• Gesture control
• Camera input
• Text prompts

And these can be freely combined. You can use gesture motions in front of the camera while typing to adjust the style — the model handles both simultaneously.

This is entirely different from traditional video models that only process prompts. A text prompt is inherently a “delayed batch-processing” interface — you write commands, submit them, wait, and get a result. But inputs like camera feeds, gestures, and video streams are continuous signals, which the model must interpret and respond to frame by frame. The engineering challenge lies not in “understanding more input types,” but in “performing a full inference cycle every few dozen milliseconds per frame.”

Three Categories, Five Core Capabilities

On top of this unified architecture, X2.0 offers five major capabilities plus one free mode:

| Capability | Code Name | Use Case | |-------------|------------|----------| | Real-time Character Swap | CharX | Switch hosts or characters in live streams | | Real-time Outfit Change | ClothX | Virtual try-on, clothing swaps | | Real-time Style Change | VibeX | Change entire scene and subject aesthetics in real time | | Touch Interaction | MoX | Drag, tap, and manipulate on-screen elements | | Dimensional Interaction | DimX | AR scenes where virtual characters respond to gestures | | Free Mode | Free | Supports custom prompts |

Each function alone isn’t entirely new, but combining them all in a single real-time responsive model makes it a different beast.

Pricing: One-Twelfth of Decart’s Rates

The boldest move this release might actually be its pricing.

Let’s look at the benchmark first. Israel-based Decart, the fastest overseas competitor, finished a $300 million funding round this May with a $4 billion valuation. Its Lucy model series is now in live testing for streaming and virtual try-on. Decart’s real-time video editing API costs $0.02 per second (720p) — that’s $72 per hour.

Xmax X2.0 costs:

• $6/hour overseas (about 1/12 of Decart’s price)
• 20 RMB/hour domestically

Let’s take a concrete example: an e-commerce shopper spends 3 minutes virtually trying on clothes. At Decart’s rate, that’s $3.6 — too expensive for most retailers. Cut that by over 90%, and mass deployment becomes viable.

The adoption barrier for real-time interactive video isn’t a matter of capability, but a matter of unit economics. X2.0’s pricing basically breaks through the cost ceiling.

Real-World Applications: No Longer Just a Demo

Live Streaming and MCN

Over the past few years, live streaming platforms have been looking for ways to deepen audience interaction. Kuaishou tried comment-based games; Douyin experimented with interactive props — but none escaped the “comment section feedback” model. The key limitation: traditional live streams are one-way, non-editable video feeds. Viewers send comments; the host reacts later — that’s an entire feedback loop.

By combining CharX + ClothX + VibeX + Free, the live room becomes a canvas viewers can directly manipulate. Type a message, the host’s outfit changes. Type another, the background becomes cyberpunk. This shifts interaction from text feedback to visual feedback — a structural leap.

Virtual Try-On for E-Commerce

ClothX’s use case is self-explanatory: you open your phone camera, point it at yourself, and your clothing changes in real time — while your body shape, motion, and lighting all remain intact.

This is completely different from today’s so-called “virtual try-ons,” which either overlay clothing on a model photo or generate static composite images. ClothX shows the live dynamic effect — “this clothing, on me, right now.” Turn around, raise your hand — the outfit moves with you.

Virtual Companions and AR

DimX aims to bridge the gap between virtual content and physical space. Make a gesture on your desk, and a virtual character appears out of thin air, perceives your position and movement, and interacts with you.

It sounds like Apple Vision Pro’s experience — except X2.0 only needs a phone camera.

Team Background

Founder Jiaxin Shi earned her bachelor’s and PhD in Computer Science from Tsinghua University and studied at the Tsinghua KEG Lab. The core team members mostly come from Tsinghua’s KEG and HCI Labs, with talent from Huawei’s “Genius Youth” program, ByteDance, and Kuaishou.

KEG is one of China’s top AI labs — Zhipu AI also originated there. HCI Lab is the country’s leader in human-computer interaction research. Combining algorithms + interaction expertise to pursue real-time interactive video is a sound path.

A Quick Judgment

The first half of 2026 will be a key window for real-time interactive video.

On one hand, Decart has large-scale funding and momentum abroad, while domestically, Xmax remains alone in the field. On the other, the quality arms race has hit diminishing returns — top models keep scaling parameters and data, but users can hardly tell the difference anymore. Among differentiating paths, “interactivity” remains one of the few that can unlock new scenarios.

Opening the API for X2.0 is also significant. In the X1 era, Xmax built its own X-cam app for users; now, X2.0 packages its capabilities for developers. This signals a shift from product to platform — the team believes the potential application space is too vast to develop alone, so they’re laying the foundation for others to build upon.

In the short term, virtual try-on and interactive live streaming will likely scale up fastest, thanks to clear ROI. In the longer run, virtual companionship and AR have greater imaginative potential but will require a broader supporting ecosystem.

OpenAI Hub is also integrating Xmax X2.0’s API, allowing developers to benchmark and test across leading video models using a unified key. As for what kinds of products will eventually emerge under this new “video-as-interface” paradigm, no one can yet say. But with the infrastructure now in place — the real game has just begun.

References

• Had a Date with My "Anime Husband"! The World’s First Playable Real-Time Interactive Model, Xmax X2.0 Released - Zhihu: Detailed report by QuantumBits on the X2.0 release