Wan vs HunyuanVideo

License Geography Decides It

Apache-2.0 lets anyone use, modify and redistribute Wan 2.2 commercially, anywhere, with an explicit patent grant and no user-count threshold. That makes it the safe default for a product company, a studio or a public institution that wants a video model it can run, fine-tune and ship without negotiating.

The Tencent Hunyuan Community License is narrower in two ways. First, its "Territory" is defined as "the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea", and the license states that it "does not apply" in those places. A company or user there has no license to use HunyuanVideo at all. Second, if a licensee's products had more than 100 million monthly active users in the month before a Hunyuan release, it must ask Tencent for a license, which Tencent may grant "in its sole discretion".

The geographic exclusion is a growing pattern among Chinese open-weight video models. MiniMax's H3 weights, released in August 2026, exclude the US as well as the EU, UK and South Korea. The licenses give no reason for the carve-outs, but the excluded markets include the EU, where the EU AI Act's transparency duties for generated media have applied since August 2026. For buyers, it means "open weights" no longer implies "usable here", and the license page needs reading before the model card.

Hardware and Running Costs

HunyuanVideo 1.5 is the lighter model to run. At 8.3 billion parameters, Tencent says it needs a minimum of 14 GB of GPU memory with model offloading, which brings it within reach of a single consumer card. The original HunyuanVideo from December 2024, with more than 13 billion parameters, recommended 80 GB.

Wan 2.2 splits the difference by size. Its dense TI2V-5B model runs at 720p and 24 fps on a 24 GB card such as an RTX 4090, producing a 720p clip in under nine minutes by Alibaba's figures. The flagship T2V-A14B and I2V-A14B models are mixtures of experts, with a high-noise expert for early denoising and a low-noise expert for detail, 27 billion parameters in total and 14 billion active per step. On a single GPU they need at least 80 GB. Both families build on the diffusion transformer design behind most current video models.

Capabilities and Ecosystem

Wan has the broader open toolkit. Beyond text-to-video and image-to-video, the 2.x line added first-and-last-frame interpolation (Wan 2.1 FLF2V), an all-in-one creation and editing model (VACE), speech-to-video (S2V-14B, August 2025) and character animation (Animate-14B, September 2025), all under Apache-2.0. That range, and the permissive license, have made Wan a common base for community fine-tunes and tools.

HunyuanVideo 1.5 is the more recent and more efficient model. It handles text-to-video and image-to-video at 480p and 720p, with a built-in super-resolution stage to 1080p and clips of 121 frames, and Tencent added a step-distilled image-to-video model in December 2025 for faster generation. Earlier Hunyuan releases covered image-to-video, customized subjects (HunyuanCustom) and talking avatars.

The top of 2026's leaderboards belongs to closed, audio-native models. Wan 3.0, Alibaba's closed successor, is first on Artificial Analysis's text-to-video board with audio, generates 30 seconds in a single pass at up to 1080p with dialogue, music and effects, and costs $0.05 to $0.20 a second on Alibaba Cloud Model Studio. No HunyuanVideo model is listed on either board (see Wan vs Seedance).

Openness Is Ending at the Frontier

Both companies have stopped releasing their best video models openly. Alibaba has shipped Wan 2.5, 2.6, 2.7 and 3.0 as API-only models, and the Wan-AI organization on Hugging Face still tops out at Wan 2.2. No HunyuanVideo successor appears on Tencent's Hugging Face or GitHub pages as of October 2026. Teams that need commercially usable open weights in 2026 should also look at Lightricks' LTX-2.5, which is free under its community license below $10 million in revenue (see LTX vs Wan). The practical consequence is that the strongest Wan and Hunyuan models anyone can download are nearly a year old or more, and teams building on them should plan for that gap.