Seedance 2.5 lands on Vercel AI Gateway: 30-second video, one call
ByteDance's audio-video model is now one AI SDK call, while MiniMax H3's open weights and Mistral's 3B Shieldstral let you self-host the whole media stack.
Copy markdown
30-second video from a single generateVideo call
Seedance 2.5 is live on Vercel AI Gateway as `bytedance/seedance-2.5` — up to 30 seconds of video with built-in audio in one pass (480p or 720p), called straight from the AI SDK's generateVideo with no separate ByteDance account. Pass a prompt plus up to 30 image, 10 clip and 10 audio references to compose multi-shot scenes in a single generation.
MiniMax H3: run an omni video model on your own GPU
The 33.1B model behind Hailuo shipped open weights on Hugging Face with native ComfyUI nodes and six workflow templates, generating up to 15-second clips (2K via upscaler) with native stereo audio across 11 languages. It's a local, no-metering path to audio-plus-video — though its community license reportedly restricts commercial use in the US and EU, so check terms before you ship.
Shieldstral: a 3B moderation model you steer in plain English
Mistral open-sourced Shieldstral under Apache 2.0 — a 3B multimodal safety classifier that runs on a single 16GB GPU and frames every check as a yes/no question, so you write the policy at inference instead of retraining. It matches or beats open guard models up to 7x its size on text, refusal detection and multimodal benchmarks, giving self-hosted apps a guardrail with no API bill.
Elsewhere: Vercel container images can now go public
Vercel Container Registry repositories can now be switched to public for read-only pulls — a quick way to share a prebuilt agent or dev image without handing out credentials.