Video
7 open-weight video models worth running at home in November 2026

Key takeaways
Wan 2.2 is still the easiest model to start with. It's Apache 2.0, ComfyUI says its 5B version should fit on 8GB of VRAM with offloading, and Alibaba's Wan-AI page on Hugging Face lists no Wan 2.5 or later weights.
Pika released H3 Flex in November, but it's a set of adapters on top of MiniMax H3, so it carries MiniMax's license, which excludes the EU, the UK, South Korea, and the US.
Kandinsky 6.0 had the busiest fall. It arrived on October 6 under the MIT license with sound included, gained first-and-last-frame input on October 25, and got a speed update on November 8.
Sound is now normal in open video. Kandinsky 6.0, LTX-2.5, MiniMax H3, and Pika H3 Flex all generate audio with the picture, and only Kandinsky 6.0 is under a license with no revenue cap or regional limit.
Makers' memory figures describe their reference setups, so check each model's low-memory options before ruling out a card.
Which open-weight video model should you run at home?
Start with Wan 2.2 if you have one consumer card and want the fewest surprises, and move to Kandinsky 6.0 when you want sound under a fully open license. Both run in ComfyUI, the free node-based app for local generation, and both can fit on a 16GB card.
This list is built for one person with one GPU, so it ranks models on how realistic they are to run at home and how freely you can use what they make. For a broader ranking, see our earlier roundup of open-weight video models. This one is a separate pass that adds what shipped since summer and sorts by home hardware and license first.
How we ranked them
Each model got the same four checks. These were the hardware its maker says it needs (along with any low-memory mode), what it generates (resolution, clip length, and sound), its license (including revenue caps and excluded regions), and how well ComfyUI supports it. Figures come from each maker's model card, repository, or license file as of November 10, 2026.
If you're unsure what your card can handle, our VRAM guide explains how offloading trades memory for speed.
What's new since June
Pika H3 Flex arrived in early November, built on MiniMax H3.
Kandinsky Lab open-sourced Kandinsky 6.0 on October 6, then added first-and-last-frame, audio-file, and video-file input on October 25.
Alibaba shipped Wan-Animate-2 in August, a character animation model, but its main video line is still Wan 2.2.
Sand AI released MAGI-1.1 24B weights in June.
LTX-2.5 and MiniMax H3 are both August releases, so they're still the current versions from their makers.
The seven models at a glance
Model | Maker | Generates sound | Maker's hardware note | License |
|---|---|---|---|---|
Wan 2.2 | Alibaba | No | 24GB (5B), 80GB (14B) | Apache 2.0 |
Kandinsky 6.0 | Kandinsky Lab | Yes, 44 kHz | Benchmarked on 16GB to 80GB cards | MIT |
LTX-2.5 | Lightricks | Yes | 32GB+ minimum | LTX-2.x Community |
HunyuanVideo 1.5 | Tencent | No | 14GB with offloading | Tencent Hunyuan Community |
Pika H3 Flex | Pika | Yes, 32 kHz stereo | Tested on an 80GB H100 | MiniMax H3 Community |
MiniMax H3 | MiniMax | Yes, 32 kHz stereo | Not published | MiniMax H3 Community |
MAGI-1.1 | Sand AI | Not stated | 24GB (4.5B) | Apache 2.0 |
1. Wan 2.2: best starting point
Alibaba released the Wan 2.2 weights on July 28, 2025, and they are still the newest general-purpose Wan models you can download.
Strengths: The 5B model handles text-to-video and image-to-video at 720p and 24 fps, and the repository says the 5B model can make a five-second 720p clip "in under 9 minutes on a single consumer-grade GPU." ComfyUI's Wan 2.2 guide says the 5B version "should fit well on 8GB vram" with native offloading, and it ships workflows for 5B, 14B text-to-video, 14B image-to-video, and first-and-last-frame. The license is Apache 2.0. Wan-Animate-2, released in August, adds a 14B character animation model under the same license.
Limits: The repository's own commands call for at least 24GB for the 5B model and 80GB for the 14B models. Wan 2.2's text-to-video and image-to-video models make silent clips, and Wan-Animate-2's model card lists ComfyUI support as a to-do.
Best for: A first local video setup on a mid-range card.
2. Kandinsky 6.0: best fully open model with sound
Kandinsky 6.0 comes in a 3B Lite size and a 29B Pro size, both MIT-licensed.
Strengths: The repository says both models make "5-second video clips with synchronized 44 kHz audio, including lip-sync." The base output is 480 by 864, and a super-resolution plug-in raises it to 1920 by 1080. Its speed table includes a 16GB RTX 5060 Ti, where Lite Distill at standard definition takes about 174 seconds for a clip. The November 8 SPEED update speeds up the Pro model by 14%, and the repository lists support in ComfyUI, Diffusers, vLLM-Omni, SGLang, and FastVideo.
Limits: Clips stop at five seconds. Pro at Full HD takes about 1,097 seconds on an RTX 4090 and 3,571 seconds on the 16GB card.
Best for: Short clips with dialogue under a license with no revenue cap.
3. LTX-2.5: best for sound at higher resolution
Lightricks' LTX-2.5 landed in August 2026 and is the version its repository recommends.
Strengths: The model card describes an open world model "built for local execution and fine-tuning" and lists native multishot generation, which can "generate connected scenes in a single pass." It also lists text-to-video, image-to-video, video-to-video, and audio modes, and examples reach 3840 by 2176 with spatial upscaling.
Limits: Lightricks' system requirements call for "a minimum 32GB+ VRAM." The repository suggests fp8 quantization and CPU or disk offloading for smaller cards, but gives no figure for how small. The model license is free for commercial use only under $10 million in annual revenue.
Best for: Multi-shot clips with sound, on a 32GB card or rented hardware.
4. HunyuanVideo 1.5: best for 720p on a 16GB card
Tencent's HunyuanVideo 1.5 dates to November 2025, and its model card has no 2026 news entries.
Strengths: The model card lists a minimum of "14 GB (with model offloading enabled)," 480p and 720p generation, and a super-resolution step to 1080p. Distilled versions cut the step count, and an 8 or 12 step version is recommended for image-to-video at 480p.
Limits: Its model card lists no audio output. The Tencent Hunyuan Community License covers "the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea," and services above 100 million monthly active users must request a license from Tencent.
Best for: 720p clips on a 16GB card, for people outside the excluded regions.
5. Pika H3 Flex: newest release, with strings attached
Pika's H3 Flex is the freshest entry on this list, though it depends on MiniMax H3. The Hugging Face repository describes LoRA adapters that apply to MiniMax H3, plus Diffusers-format H3 base components that run to about 210GB.
Strengths: It generates 5 to 15 seconds at 24 fps in native 480p or 768p, with 32 kHz stereo audio. Modes include text-to-video, first-and-last-frame, and a reference mode that takes up to nine images, three videos, and three audio clips. The repository lists ComfyUI workflow templates.
Limits: The 1080p to 4K upsampler and audio stem separation run only on Pika's hosted API. The card gives no VRAM figure, only that sample requests passed on an 80GB H100 and that large ones may need CPU offloading and plenty of system RAM. It uses the MiniMax H3 Community License, so MiniMax's exclusions apply.
Best for: Reference-driven clips with sound, for readers who can use the MiniMax license and have workstation-class hardware.
6. MiniMax H3: best base for reference-driven clips
MiniMax H3 is the 33B model underneath Pika's release, and its weights are on Hugging Face.
Strengths: It generates 4 to 15 seconds at 24 fps, with a default of 768 pixels on the short side and 32 kHz stereo sound in 11 dialogue languages, according to its model card. ComfyUI has an official tutorial and templates for reference-to-video and text-to-video.
Limits: The license says "Excluded Territories" means "the European Union, the United Kingdom, the Republic of Korea and the United States of America," and it requires written authorization above $20 million in yearly revenue. The card gives no VRAM figure, its SGLang example uses four GPUs, and the 2K module isn't open-sourced.
Best for: Multi-reference projects, outside the excluded regions.
7. MAGI-1.1: best for stretching a clip
Sand AI's MAGI-1 is Apache 2.0, and the 24B MAGI-1.1 weights followed on June 17, 2026.
Strengths: It supports text-to-video, image-to-video, and video-to-video, so you can continue an existing clip. The repository says "any machine with at least 24GB of GPU memory is sufficient" for the 4.5B model, and ComfyUI custom nodes have existed since May 2025.
Limits: The 24B models call for eight H100 or H800 cards, or four of those or eight RTX 4090s with the distilled fp8 build. The repository doesn't describe audio generation.
Best for: Extending clips on a 24GB card with the 4.5B model.
Other models you may still see
SkyReels V3 from Skywork has been out since January 2026, with separate reference-to-video, extension, and talking-avatar models. Its repository recommends a low-VRAM mode for cards under about 24GB and doesn't mention ComfyUI. Wan 2.7 and other later Wan versions turn up in search results, but Alibaba's Wan-AI page on Hugging Face lists no weights for them. If a site offers a Wan 2.7 download, check it against that page first.
How the licenses split
Apache 2.0 and MIT, which cover Wan 2.2, Kandinsky 6.0, and MAGI-1.1, have no revenue cap or regional limit. The community licenses add conditions. LTX-2.5 has a $10 million revenue line, HunyuanVideo 1.5 excludes the EU, the UK, and South Korea, and MiniMax H3 and Pika H3 Flex exclude those regions plus the US. If you plan to train add-on files, our LoRA explainer covers how they relate to the base model.
Open-weight video models FAQ
What is the best open-weight video model for a home GPU?
Wan 2.2 for most setups, since it's Apache 2.0 and ComfyUI says its 5B version should fit on 8GB with offloading. Kandinsky 6.0 is the pick if you want sound under an MIT license.
Is Pika H3 Flex a full model?
No. Its repository holds adapters that apply to MiniMax H3, plus converted base components, and it's under the MiniMax H3 Community License. The upsampler and audio stem tools stay on Pika's hosted API.
Is Wan 2.5 or Wan 2.7 open source?
Not as far as Alibaba's Wan-AI page on Hugging Face shows. It lists Wan 2.2 models and add-ons such as Wan-Animate-2, with no weights for later versions.
Which open video models generate audio?
Kandinsky 6.0, LTX-2.5, MiniMax H3, and Pika H3 Flex generate sound with the picture. Wan 2.2's text-to-video and image-to-video models and HunyuanVideo 1.5 don't list audio output.
Can I use these models commercially?
Wan 2.2, Kandinsky 6.0, and MAGI-1.1 can be used commercially under Apache 2.0 or MIT. The others come with revenue thresholds or excluded regions, so read the license file for your situation.
