On the last day of July 2026, two of China's largest AI labs shipped new video models within hours of each other. One kept its weights behind a closed door. The other promised to hand them over.
For anyone making films, ads, or microdramas (短剧), the difference is not technical trivia. It decides who controls the footage you generate, where it is processed, and what you pay per second.
The two launches
ByteDance pushed Seedance 2.5 out to Chinese consumers through Jimeng AI (即梦), the creation app born from the Jianying (CapCut China) team, and to the Pro tier of its Doubao assistant, with a developer API following on its Volcano Engine (BytePlus) platform. Seedance is the video arm of ByteDance's Seed large model (大模型) family.
The same day, Shanghai's MiniMax released H3 — the newest generation of its Hailuo (海螺) video line — through the Hailuo AI consumer app and a developer API. "Hailuo 3.0" is a community nickname; MiniMax's own name for the model is H3.
Both target the same customers: studios, agencies, and the microdrama (短剧) producers pumping out 30-second episodes for China's short-video platforms.
Seedance 2.5: one long take
Seedance 2.5's headline trick is a continuous 30-second clip generated in a single inference pass, rather than several short clips stitched at the edit points where faces and lighting usually drift. ByteDance credits a sparse diffusion transformer that samples across the full clip instead of a sliding window.
The consumer version accepts up to 50 reference assets per generation — 30 images, 10 videos, 10 audio clips — so a brand can lock a product's look, a character's face, and a scene's style at once. Audio is generated in the same latent space as the video, keeping dialogue and ambient sound in sync without a separate step. ByteDance also claims native 4K rendering and 10-bit color.
The catch: that 4K claim has not been confirmed by independent third-party benchmarks, and a 180-second beta mode spotted in the app is not part of the official feature set.
Hailuo's H3: open by design
H3 takes a different bet. MiniMax describes it as a single unified model — H3-Omni — that ingests text, image, video, and audio as one context and returns a finished clip with synchronized stereo sound generated in the same pass. It produces shorter clips than Seedance, about 4 to 15 seconds, at native 2K and a fixed 24 fps, and accepts up to 12 reference inputs.
MiniMax says the unified training lifted end-to-end throughput by roughly 30%. The move that got attention, though, was the license: the company said it would publish H3's weights within days, pending regulation, letting developers run the model on their own hardware with no MiniMax server in the loop.
For a microdrama (短剧) studio in Jakarta or Ho Chi Minh City, that distinction is legal, not aesthetic. Data-residency rules make sending unreleased client footage to an offshore endpoint a contract problem. A downloadable model keeps the assets inside the building.
Why the price tag is the real story
The lever Chinese labs have used against US video models all year is price, and now they are using it on each other. MiniMax listed H3 at 0.8 yuan per second for 2K video — about US$0.11 per second at roughly 7.1 RMB to the dollar — with native synchronized audio. At that rate a 15-second 2K clip lands near US$1.50–2.00, inside the impulse-test range for a regional agency.
That undercuts Western flagships on cost per finished asset, not on peak quality. ByteDance did not publish a consumer per-second price for Seedance 2.5 at launch, leaning instead on its distribution through Jimeng and Doubao.
Where this leaves creators
Two Chinese video large models (大模型) now bracket the market: one closed and productized, one open and portable. The practical questions shift from "which looks better" to "where does my footage live" and "can I fine-tune it."
For microdrama (短剧) pipelines especially, the open-weight path means a studio can bake H3 into its own render farm and iterate on style without per-second metering. The closed path means tighter integration with ByteDance's editing and distribution ecosystem — Jimeng feeds straight into Jianying and Douyin.
What readers can do now
- If you produce video commercially, run a side-by-side on your own assets: Seedance 2.5 for longer single-take sequences and brand-consistent reference locking; Hailuo H3 where data residency or local fine-tuning matters.
- If you build tooling, treat the open-weight promise as a procurement clause, not a given — verify the weights repository and licence terms before committing a pipeline.
- If you follow the market, watch price per finished second, not benchmark scores; that is where Chinese video models are pressuring Sora and Veo.
Honest limitations
Core facts (31 July 2026 same-day launches of ByteDance Seedance 2.5 via Jimeng AI/即梦 and Doubao Pro, with Volcano Engine/BytePlus API; MiniMax H3 via Hailuo AI/海螺 and developer API; Seedance 2.5 single-pass 30-second generation, up to 50 reference assets, native-4K claim unverified by third parties; H3 4–15 sec at 2K/24fps, unified H3-Omni architecture, ~30% throughput gain, stated plan to release weights pending regulation; MiniMax H3 pricing of 0.8 yuan/sec at 2K ≈ US$0.11/sec at ~7.1 RMB/USD) come from TechTimes, National Business Daily (每日经济新闻), and AI-in-Asia reporting that cites Reuters, SCMP, Bloomberg, and Caixin. Currency conversion uses ~7.1 RMB per US dollar. The "open weights" claim is MiniMax's stated intention as of launch; independent confirmation that a public weights repository shipped and its exact licence terms should be verified before reliance. Benchmark and spec figures (4K, 180-second beta) are vendor or early-media claims, not yet independently benchmarked. Analysis is current to 29 September 2026 and is not investment advice.