Key takeaways
MiniMax H3, opened in July 2026, generates four to 15 seconds with 32 kHz stereo audio and takes first frames, last frames and references; locally it renders at 768p, while 2K runs only through MiniMax's API.
LTX-2.5 from Lightricks is the open model for high resolution, with 4K output and synchronized audio, free for companies under $10 million in annual revenue.
Wan 2.2 is fully Apache 2.0 and its 5B version runs on a single RTX 4090, though it's silent and tops out at 720p; newer Wan versions are not open.
Kandinsky 6.0 Video is MIT-licensed and makes short clips with synced speech, music and effects.
Check the license before client work: Hunyuan models exclude the EU, UK and South Korea, and MiniMax H3 and LTX-2.5 add terms for larger companies.
How we ranked
These are models with downloadable weights, not hosted apps. The order weighs resolution and clip length, whether the model generates sound, image-to-video and reference support, what hardware it realistically needs and how permissive the license is for commercial work. Details come from each model's official card, repository and license file, checked November 2, 2026.
Open-weight video models at a glance
Model | Max output (local) | Audio | Hardware | License |
|---|---|---|---|---|
MiniMax H3 | 4 to 15 s at 768p, 24 fps | Yes, stereo | Multi-GPU by default; quantized builds for consumer cards | MiniMax H3 Community License |
LTX-2.5 | Up to 4K, up to 20 s | Yes | About 66 GB download; fp8 and offloading for smaller GPUs | Free under $10M revenue |
Wan 2.2 | 720p, 24 fps | No (separate speech-to-video model) | 5B runs on an RTX 4090 | Apache 2.0 |
Kandinsky 6.0 Video | About 5 s at 24 fps; upscaler to 1080p | Yes | Lite (3B) for consumer GPUs; Pro (29B) needs much more | MIT |
HunyuanVideo 1.5 | 480p and 720p; upscaler to 1080p | No | 14 GB minimum with offloading | Tencent Hunyuan Community (not EU, UK or South Korea) |
MAGI-2 Preview | 10 s; refiner to 1080p | Yes | Eight data-center GPUs | Apache 2.0 |
1. MiniMax H3: Best for long clips with stereo sound
MiniMax released the weights for H3, the model behind its Hailuo app, on Hugging Face in July 2026.
What works: The 33-billion-parameter model makes four to 15 seconds at 24 fps with 32 kHz stereo audio in the same pass. One checkpoint handles text, first frames and last frames, and a reference checkpoint takes up to nine images, three video clips and three audio clips. Community runners such as Wan2GP add quantized and accelerated variants for consumer cards.
What doesn't: Local output defaults to 768p, since the 2K stage is API-only. Products earning more than $20 million a year need MiniMax's written approval, and commercial products must display "MiniMax H3."
Best for: Short scenes with dialogue and sound on your own hardware.
2. LTX-2.5: Best for 4K output
Lightricks publishes LTX-2.5 as a 22-billion-parameter open model alongside its LTX Studio app.
What works: It generates text-to-video and image-to-video with synchronized audio, at 1024 by 1536 by default and up to 4K (3840 by 2176), with optional 2x or 4x temporal upscaling for smoother motion. The official repository includes fp8 and CPU or disk offloading options for GPUs with less memory.
What doesn't: The download is about 66 GB, and companies with $10 million or more in annual revenue need a paid commercial license.
Best for: Creators and small studios who want high-resolution output locally.
3. Wan 2.2: Best for a permissive license
Wan 2.2 is the latest open release of Alibaba's Wan video models.
What works: The A14B mixture-of-experts models handle text-to-video and image-to-video, and the 5B model does both at 720p and 24 fps on a single RTX 4090. Everything is Apache 2.0, so commercial use carries no revenue cap, and add-on models cover speech-to-video and character animation, including Animate-2 from July 2026.
What doesn't: The base models make no sound, output stops at 720p and Wan 2.5 and later versions are closed, so this is an older generation than the others here.
Best for: Commercial projects and fine-tuning, where the license matters most.
4. Kandinsky 6.0 Video: Best MIT-licensed model with audio
Kandinsky 6.0 Video arrived in October 2026 from Kandinsky Lab, part of the Russian bank Sber.
What works: Lite (3 billion parameters) and Pro (29 billion) generate video with synchronized audio, including speech with lip sync, music and effects, from text or an image. A separate super-resolution model upscales to 1080p, Diffusers supports it and the MIT license allows commercial use.
What doesn't: Clips are about five seconds long, and the Pro model needs far more memory than a typical desktop GPU.
Best for: Short talking or musical clips with a permissive license.
5. HunyuanVideo 1.5: Best for modest GPUs
HunyuanVideo 1.5 is Tencent's 8.3-billion-parameter model from November 2025, trimmed for mid-range cards.
What works: It offers 480p and 720p text-to-video and image-to-video checkpoints, a super-resolution network to 1080p and a step-distilled image-to-video model for faster runs. The official card lists a 14 GB minimum with offloading, and community runners go lower, while training and LoRA code are public for fine-tuning.
What doesn't: It generates no sound, and its license does not apply in the EU, UK or South Korea.
Best for: Image-to-video on a mid-range graphics card outside those regions.
6. MAGI-2 Preview: Best for research-grade audio-video
Sand AI's MAGI-2 Preview is a mixture-of-experts model with 114 billion parameters, 6 billion active per token.
What works: It generates 10-second clips with sound from text or an image, then refines them up to 1080p, under the Apache 2.0 license.
What doesn't: It requires eight NVIDIA Hopper data-center GPUs, which puts it out of reach for home setups, and it's still labeled a preview.
Best for: Labs and studios with cloud GPUs that want a permissive audio-video model.
Older open models such as Mochi 1, CogVideoX and Open-Sora 2.0 haven't had new releases in 2026.
How to run open video models
Start with the official repository or Diffusers, which supports Kandinsky 6.0, Wan and HunyuanVideo.
Wan2GP packages many of these models with low-memory settings for consumer GPUs.
Use quantized (fp8) weights and CPU offloading when VRAM is tight, at the cost of speed.
Generate short drafts at low resolution, then upscale the keepers.
Read the license file in each repository before using output in paid work.
Frequently asked questions
What is the best open-weight video model?
MiniMax H3 is the strongest open-weight video model for most creators in November 2026, with four to 15 seconds of video and stereo sound in one pass. LTX-2.5 reaches 4K, and Wan 2.2 has the most permissive license.
Can I run an AI video model on my own PC?
Yes. Wan 2.2's 5B model runs on an RTX 4090, HunyuanVideo 1.5 needs about 14 GB with offloading and quantized builds of MiniMax H3 and LTX-2.5 run on consumer cards through community runners.
Which open video models generate sound?
MiniMax H3, LTX-2.5, Kandinsky 6.0 Video and MAGI-2 Preview generate audio with the video. Wan 2.2 and HunyuanVideo 1.5 are silent.
Can I use open-weight video models commercially?
Usually, within each license. Wan 2.2 and MAGI-2 are Apache 2.0 and Kandinsky 6.0 is MIT. LTX-2.5 needs a paid license above $10 million in revenue, MiniMax H3 needs approval above $20 million and Hunyuan's license excludes the EU, UK and South Korea.
Are Wan 2.5 and Wan 3.0 open source?
No. Alibaba's newest open Wan weights are Wan 2.2 and its add-on models; Wan 2.5 and later versions are available only as hosted services.