Alibaba AI API

HappyHorse 1.1 Text-to-Video is a video generation model from Alibaba's ATH (Alibaba Token Hub) AI Innovation Unit, generating video and synchronized audio together in a single pass rather than adding sound afterward. It builds on HappyHorse 1.0, with Alibaba reporting improvements in motion dynamics, subject consistency, prompt adherence, visual quality, and audio generation. The model outputs clips from 3 to 15 seconds at 720p or 1080p, in 16:9, 9:16, 1:1, 4:3, or 3:4 aspect ratios, with lip-synced dialogue, ambient sound, and effects generated alongside the visuals. Following its release in June 2026, it ranked No. 2 across Artificial Analysis's Video Arena leaderboards, behind Google's Veo 3.1.

HappyHorse 1.1 Image-to-Video is a closed video generation model from Alibaba that turns a single reference image, used as the first frame, into a short video clip, guided by an optional text prompt. It's part of Alibaba's HappyHorse line, distinct from the open-weight Wan models built by Alibaba's Tongyi Lab. Output resolution goes up to 1080p, with clip lengths from 3 to 15 seconds across five aspect ratios (16:9, 9:16, 1:1, 4:3, 3:4). The model generates audio alongside video in the same pass, including lip-synced dialogue in multiple languages, rather than adding sound in a separate step. Compared to HappyHorse 1.0, Alibaba describes 1.1 as improving motion dynamics, skin texture realism, and close-up quality, though we haven't seen an independent benchmark confirming those gains.

HappyHorse 1.1 Reference-to-Video is a video generation model from Alibaba, built by the company's ATH (Alibaba Token Hub) AI Innovation Unit. Its reference-to-video (R2V) mode takes 1 to 9 reference images of characters or scenes together with a text prompt and composes them into a single video, keeping each subject's identity consistent across shots. HappyHorse is a 15-billion-parameter unified self-attention Transformer that generates video and audio in one pass rather than as separate steps. Version 1.0 first appeared anonymously in the Artificial Analysis Video Arena in April 2026 before Alibaba was named as its creator; version 1.1 followed in June 2026 with stronger motion expressiveness, tighter multi-image consistency, and improved audio-visual sync. Output runs 3 to 15 seconds at 720p or 1080p across multiple aspect ratios.

HappyHorse 1.0 Reference-to-Video is a video generation model from Alibaba's ATH (Alibaba Token Hub) AI Innovation Unit that generates new video scenes from 1 to 9 reference images plus a text prompt, keeping character identity, outfits, and visual style consistent across the output. It is one of several generation modes in the HappyHorse 1.0 family, alongside text-to-video, image-to-video, and video editing. Output ranges from 3 to 15 seconds (5 seconds by default) at 720p or 1080p, with synchronized audio generated alongside the video by the same 15-billion-parameter transformer. HappyHorse 1.0 topped the text-to-video and image-to-video rankings on the Artificial Analysis Video Arena benchmark after its April 2026 release.

HappyHorse 1.0 Image-to-Video is a video generation model from Alibaba's ATH AI Innovation Unit that animates a static image, with an optional text prompt, into a video clip while preserving the source subject. It outputs at 720p or 1080p resolution in clips from 3 to 15 seconds, and generates synchronized audio, including dialogue, ambient sound, and effects, in the same pass as the video. At launch in April 2026, it topped the Artificial Analysis Video Arena leaderboard for both text-to-video and image-to-video generation. HappyHorse 1.0 is part of a broader model family from the same unit that also includes text-to-video and subject-to-video generation, plus video editing modes.

HappyHorse 1.0 Text-to-Video is a video generation model from Alibaba that turns a text prompt into a short video clip. It's served through Together AI as one of several hosting platforms, alongside Cloudflare, fal, and others. The model generates clips from 3 to 15 seconds long at 720p or 1080p resolution, in aspect ratios including 16:9, 9:16, 1:1, 4:3, and 3:4. At launch, HappyHorse 1.0 topped the Artificial Analysis Video Arena leaderboard for both text-to-video and image-to-video, beating ByteDance's Seedance model in blind human evaluations. It first appeared on the leaderboard without a disclosed maker until Alibaba confirmed it as the model's creator.

Tongyi DeepResearch 30B A3B is an agentic large language model from Alibaba's Tongyi Lab, purpose-built for long-horizon, multi-step information-seeking and web research tasks. It uses a Mixture-of-Experts architecture with 30.5B total parameters but only 3.3B activated per token, keeping inference costs low. The model achieves state-of-the-art results across agentic research benchmarks, scoring 32.9 on Humanity's Last Exam, 43.4 on BrowseComp, 70.9 on GAIA, 75.0 on xbench-DeepSearch, and 90.6 on FRAMES — outperforming OpenAI o3 and DeepSeek-V3.1 on most of these tasks. It supports a 128K context window and two inference modes: a standard ReAct mode and a heavier iterative research mode for maximum performance. Best suited for developers building autonomous research agents, deep fact-finding pipelines, or complex multi-source synthesis workflows — especially where cost efficiency matters.