On July 31, 2026, the Chinese AI video generation market delivered two major announcements in a single day. MiniMax unveiled its H3 model — a multimodal video generation system producing clips at 2K resolution with native stereo audio — while ByteDance launched Seedance 2.5, extending single-generation video duration to 30 seconds. Together, the launches mark a turning point in the global AI video race, arriving months after OpenAI quietly discontinued its Sora product line.
Key Highlights
- MiniMax H3 generates 2K video at 24fps with natively synced dialogue, ambience, and music in a single pass
- ByteDance Seedance 2.5 produces 30-second single-take video clips — double the duration of its predecessor
- Both models launched on the same day, July 31, 2026
- MiniMax plans to release H3 model weights publicly within days
- OpenAI discontinued Sora on March 24, 2026, leaving a strategic gap in the market
MiniMax H3: 2K Video With a Full Audio Stack
MiniMax H3 is a general-purpose multimodal model designed to generate video from text, images, existing footage, and audio reference clips. Unlike earlier models that require separate audio post-processing, H3 generates dialogue, ambient sound, and music in the same pass — producing natively integrated stereo audio with no stitching required.
Clips can be 4 to 15 seconds long at 2K resolution and 24fps. Users can supply up to 12 reference files per request: 9 images, 3 reference videos, and 3 audio clips. The model also supports motion transfer — generating video that mimics the movement pattern of a reference clip applied to different content.
At $0.13 per second for 2K output, MiniMax claims its pricing is less than one-third of competing products. A 15-second 2K clip costs approximately $1.95 at base rate. The company has announced plans to release H3 model weights publicly within days, joining the growing wave of open-weight AI models from Chinese labs.
MiniMax's Hong Kong-listed shares rose between 13% and 17.4% intraday following the announcement.
ByteDance Seedance 2.5: The 30-Second Barrier
ByteDance's Seedance 2.5 doubles the maximum single-generation duration from 15 seconds to 30 seconds — a significant leap that enables longer narrative sequences without manual stitching. The model accepts up to 30 reference images, 10 videos, and 10 audio clips in a single request, giving creators fine-grained control over the visual and audio style of the output.
New editing capabilities include timestamp-based controls — allowing users to target specific moments in a clip — and green-screen editing for compositing workflows. ByteDance targets education, industrial, and entertainment use cases alongside advertising and marketing.
Seedance 2.5 is rolling out on Jimeng AI and the Doubao Pro app, with API access expected on Volcano Engine's Ark platform.
Filling the Sora Vacuum
The simultaneous launches arrive at a strategically important moment. OpenAI discontinued Sora on March 24, 2026, with the web and app versions going offline on April 26 and API services set to fully terminate on September 24, 2026. The exit by the company that first captured global attention with AI video has left a significant vacuum — one Chinese labs are clearly moving to fill.
The launches also follow ByteDance's Seedance 2.0, which reached over $147 million in monthly revenue as of June 2026, demonstrating strong commercial demand for high-quality AI video generation.
Impact on the AI Video Market
The simultaneous releases reflect a broader competitive dynamic: Chinese companies are now shipping frontier video AI at commercial scale, with open weights and aggressive pricing designed to attract global developers.
For developers and businesses in MENA markets, both models offer direct API access at prices well below earlier generation tools. MiniMax H3's open-weight release will enable on-premise deployment — an important consideration for organizations with data sovereignty requirements.
What's Next
MiniMax expects to release H3 weights on Hugging Face within days. ByteDance has not announced open-source plans for Seedance 2.5, though API access on Volcano Engine is imminent. Industry observers expect both models to be integrated into major cloud platforms by Q3 2026, further accelerating global adoption of AI-generated video for commercial production.
Source: MiniMax Blog