A marketing lead types a single sentence into a box — "a 20-second reel of a coffee shop opening at dawn, upbeat" — and gets back a clip with camera moves, music, and a voiceover. No timeline, no editor, no render farm.
That is the bet behind Wondershare (万兴科技, 300624.SZ), the Changsha-based maker of Filmora and other creative software: that the next wave of users does not want tools, they want results.
China's first audio-video model
On 30 January 2024, Wondershare launched Tianmu (天幕) at Malanshan in Changsha. It was billed as China's first audio-video multimedia large model (音视频多媒体大模型) and the first in Hunan province to clear the algorithm-filing requirement. A public beta followed on 28 April 2024.
The name fits the ambition: "Tianmu" means a sky used as a stage curtain. The model was built to handle the full chain — text-to-video, image-to-video, video continuation, text-to-image, and audio generation including music, sound effects, and voice — rather than a single trick.
Version 2.0, with Huawei inside
On 22 June 2025, at Huawei's HDC 2025 Pangu forum, Wondershare released Tianmu 2.0, built on Huawei Cloud's (华为云) Pangu (盘古) multimodal model. The company says average performance rose about 90% over 1.0, with gains in professional camera-movement control and layered stereo-sound generation. It was trained on 10-billion-scale data plus 1.5 billion user-behavior records from Wondershare's two decades in creative software.
At the same event, Wondershare and Huawei Cloud opened an "AI video model lab" to keep co-developing the stack.
The benchmark bragging rights
Wondershare points to external leaderboards for credibility. On VBench-2.0, its overseas product ToMovie 2.0 ranked in the global top three, and took first place worldwide in two sub-metrics: camera motion and motion reasonableness. The base model also placed inside China's top four on the SuperCLUE global board, with the company claiming it beat OpenAI's Sora on that particular measure.
Those are vendor-cited rankings, so read them as "we are in the conversation," not "we won."
How it is actually used
Tianmu is not a research demo. Wondershare says cumulative model calls passed 500 million in the first half of 2025, and that AI-native application revenue crossed ¥60 million (≈ US$8.5M / HK$66M) over the same period. The per-clip cost on the consumer app reportedly runs as low as about ¥0.25 (≈ US$0.04 / HK$0.27).
The surfaces:
- Wondershare Filmora (万兴喵影), which in its 2026 edition claims an end-to-end flow from AI素材 generation to fine editing inside one general-purpose video app.
- ToMovie (tomoviee.cn), a web creation plaza for professional users, now with a mobile app.
- 超媒Agent (Chaomei Agent), a "one sentence generates a video" assistant with multi-turn dialogue editing.
Wondershare has also connected some overseas products to Google's Veo 3 and Nano Banana Pro — a reminder that Chinese creative software increasingly mixes homegrown and foreign models.
Why a tools company built its own model
Wondershare's chairman Wu Taibing (吴太兵) frames this as "creative equality" (创意平权) — lowering the bar so anyone can produce. But the business logic is sharper: if the model is the product, a software license becomes a usage-based subscription, and the company stops renting features and starts selling output.
The risk is the same for every model-builder: training and inference costs. Wondershare's answer is vertical focus — video, not everything — and a "dual-engine" setup where self-built models make up the majority of calls, with third-party models filling gaps.
Why this matters beyond China
Video generation is moving from research demo to packaged software. When a consumer app maker ships a multimedia large model (大模型) inside its product, the technology stops being a curiosity and becomes infrastructure that small creators depend on. The open question is whether quality keeps pace with the marketing — and whether platforms give creators transparent labels on AI-generated clips.
Honest limitations
The call counts, revenue figure, and ¥0.25-per-clip cost come from Wondershare's own disclosures and Baidu Baike, which aggregate company claims; none are independently audited. The "beat Sora" claim refers to one vendor-cited leaderboard and one metric, not a broad consensus. Tianmu is a video-vertical model, so it is not comparable to frontier general models on general reasoning or image quality. And the ¥0.25 figure is a promotional low end — real multi-shot, music-backed clips cost more. Wondershare's heavier reliance on its own models also raises the question of whether it can keep pace with better-funded rivals shipping new video models every few months.
What readers can do now
- If you edit video, try Filmora's 2026 AI flow on a real project and time the difference between AI素材 generation plus editing versus your usual pipeline.
- If you produce at volume, open ToMovie (tomoviee.cn) and price a batch of short clips at the published per-clip cost against a human editor.
- If you follow the China AI market, watch whether Wondershare's "model as product" subscription revenue keeps climbing past the ¥60 million (≈ US$8.5M / HK$66M) half-year mark — that is the proof the bet is paying off.