国产AI视频模型又卷起来了：Seedance 2.5和MiniMax H3谁更值得用

TL;DR - Two Chinese video-generation models, ByteDance's Seedance 2.5 and MiniMax H3, launched on the same day at the end of July, splitting the market between long-form narrative generation and cheap, controllable commercial editing.

• Seedance 2.5 (live in Jimeng/Doubao Pro; API pending) extends single-shot generation from 15s to 30s with multi-round continuation that preserves character, scene, style, and voice; it accepts up to 30 images + 10 videos + 10 audio clips with timestamp-level targeted editing. API price rose ~50% over 2.0 while resolution was capped at 720P (no 4K), reflecting higher compute cost.
• MiniMax H3 is a unified omni-modal model (text/image/video/audio understanding) at up to 2K resolution and 4–15s clips, already API-available domestically and internationally, accepting 9 images + 3 videos + 3 audio references per request.
• Pricing gap is the headline: H3 at ¥0.8/s (2K) and ¥0.5/s (768P) vs. Seedance's ¥10/s — a 15s 2K H3 clip lists at $1.95. MiniMax also says it will open-source H3 weights.
• H3 targets commercial output: strengthened text/logo/packaging rendering without cross-frame distortion, plus video-to-video motion transfer; it currently ranks #1 on Artificial Analysis' video-editing arena (by 9 points, benchmarked against Seedance 2.0, not 2.5).