Every guide, comparison and workflow we’ve published on Alibaba.
10 articles
Happy Horse 1.0 Reference to Video is reference-to-video. Upload the stills that must appear. Do not spend the whole budget because the slider goes that far.
Happy Horse 1.0 Text to Video is one catalog generate. Use it because this job matches this row.
Happy Horse 1.0 Video Edit edits footage you already have. A new generate is a different identity.
Qwen Image 2 Edit is an edit. Point at the object. A new text-to-image throws away the still you approved.
Qwen Z Image is the stills row. Animate after approval. Do not generate the pack inside the video model.
Wan Video 2.5 Image to Video is image-to-video. If the label is wrong, every second is waste.
Alibaba quotes ~38 seconds per 1080p HappyHorse 1.1 clip on one H100. What a hardware-conditioned figure predicts about a hosted queue, and what it can't.
No Hugging Face checkpoint, no repo and no ComfyUI node exists for Wan 3.0. How to write a self-host plan that survives a weights rumour turning out false.
Alibaba's Wan 3.0 beta claims 30-second single takes and video generated from PDFs, decks, and spreadsheets. A skeptical look, and what Wan 2.7 does today.
A research-backed breakdown of Qwen 3.5 - Alibaba's 397B-A17B MoE that crossed 942M Hugging Face downloads, its native multimodal stack, coding benchmarks vs DeepSeek V4 and Llama 4, and how creators can route through it on OpenRouter.
