The best open-weight video models in November 2026

Key takeaways

• Wan 2.2 is the place to start for local video. It's Apache 2.0, ComfyUI runs its 5B model on 8GB of VRAM with offloading, and Alibaba hasn't released Wan 2.5 or anything later as open weights.

• Three open models now generate sound with the picture: LTX-2.5, Kandinsky 6.0, and MiniMax H3. Only Kandinsky 6.0 is under a fully open license (MIT).

• "Open weights" doesn't always mean open license. LTX-2.5 is free only under $10 million in annual revenue, and the HunyuanVideo 1.5 and MiniMax H3 licenses exclude the European Union, the UK, and South Korea. MiniMax's also excludes the United States.

• Makers' VRAM figures run high. Lightricks asks for 32GB or more for LTX-2.5 and Alibaba lists 24GB for Wan 2.2 5B, but offloading in ComfyUI fits several of these models on 16GB cards or smaller, more slowly.

• MAGI-1 and SkyReels V3 are specialists: MAGI-1 extends clips chunk by chunk, and SkyReels V3 handles reference images, clip extension, and talking avatars.

Which open-weight video model is best?

For most people running video on their own graphics card, start with Wan 2.2. It's Apache 2.0, ComfyUI has native workflows for its main modes, and its 5B model fits small cards. If you need sound or higher resolution and have 32GB of VRAM, LTX-2.5 is the next one to try.

An open-weight model is one whose trained files you can download and run yourself. The license on those files still decides who can use them commercially and where, so this list separates true open licenses (Apache 2.0 and MIT) from community licenses with revenue caps, regional exclusions, or content rules.

How we ranked them

We ranked the seven on what each one generates (resolution, length, and audio), the hardware its maker says it needs, how freely its license lets you use it, and its support in ComfyUI, the free node-based app for running models locally. Figures come from each maker's model card, GitHub repository, or license file as of early November 2026.

The seven models at a glance

1. Wan 2.2: best overall

Alibaba's Wan team released Wan 2.2 in July 2025, and it's still the newest Wan video family with downloadable weights. Wan 2.5 and later run only on hosted services.

Strengths: The family covers text-to-video and image-to-video at 480p and 720p with two 14B models, each built from two experts, plus a 5B model that does both modes at 720p and 24 fps. Everything is Apache 2.0, and the model card says the team claims no rights over what you generate. ComfyUI's Wan 2.2 guide says the 5B model "should fit well on 8GB vram" with native offloading, and ships workflows for 5B, 14B text-to-video, 14B image-to-video, and first-and-last-frame. Add-ons keep arriving, most recently Wan-Animate-2-14B in August 2026, which animates a character image from a driving video.

Limits: Wan 2.2 makes silent clips. Alibaba's reference code needs at least 24GB for the 5B model and 80GB for the 14B models, and the model card puts a 5-second 720p clip from the 5B model at "under 9 minutes" on a consumer GPU.

Best for: Anyone starting with local video, and anyone who needs a license with no revenue cap or content rules.

2. LTX-2.5: best for sound and 4K

LTX-2.5 is a 22B model from Lightricks, released in August 2026, that generates synchronized video and audio in one pass.

Strengths: Its model card lists text-to-video, image-to-video, audio-to-video, video-to-video, and combined audio-and-video modes, and the GitHub repo adds keyframe interpolation, a retake mode, and 4K output at 3840 by 2176. ComfyUI ships official LTX-2.5 workflow templates. Lightricks' examples run 121 frames at 24 fps, about five seconds.

Limits: Lightricks' system requirements call for an NVIDIA GPU with "a minimum 32GB+ VRAM," 32GB of system RAM, and 100GB of free storage, and recommend an A100 or H100. The LTX-2.x Community License allows free commercial use only for companies under $10 million in annual revenue, and it makes Lightricks' Acceptable Use Policy binding, which bans sexually explicit content and gratuitous gore.

Best for: Clips that need dialogue or sound effects, on a 32GB card or rented hardware.

3. Kandinsky 6.0: best fully open model with audio

Kandinsky Lab open-sourced Kandinsky 6.0 on October 6, 2026, in a 3B Lite size and a 29B Pro size.

Strengths: Both sizes generate "5-second video clips with synchronized 44 kHz audio," including lip-sync, from text or from an image. The base output is 480 by 864, and a separate super-resolution model raises it to 1920 by 1080. It's MIT-licensed, and an October 25 update added first-and-last-frame input, video driven by an audio file, and audio driven by a video file. The GitHub repo's speed table includes a 16GB RTX 5060 Ti, with Lite fully on the card and Pro using block offloading.

Limits: Clips stop at five seconds. Pro is slow on consumer hardware: about 18 minutes (1,096.7 seconds) for a Full HD clip on an RTX 4090. ComfyUI support comes through community nodes in the ComfyUI registry.

Best for: Short clips with sound under a license with no strings, including on a 16GB card.

4. HunyuanVideo 1.5: best for 16GB cards

At 8.3B parameters, HunyuanVideo 1.5 is far smaller than Tencent's 13B original, and it came out in November 2025.

Strengths: The model card lists text-to-video and image-to-video at 480p and 720p, with super-resolution models that step up to 1080p. It needs "14 GB (with model offloading enabled)," and the default clip is 121 frames. Tencent publishes an official ComfyUI guide and step-distilled variants for faster runs.

Limits: The Tencent Hunyuan Community License defines its territory as worldwide "excluding the territory of the European Union, United Kingdom and South Korea." Services with more than 100 million monthly active users need a separate license, and its use policy requires labeling machine-generated content you publish. Clips are silent.

Best for: A 16GB card, outside the regions the license excludes.

5. MiniMax H3: best for reference-heavy clips

H3 started as a MiniMax API in July 2026, and the weights for the 33B model followed in August.

Strengths: H3 generates 4 to 15 seconds at 24 fps, with 768 pixels on the short side by default and 32 kHz stereo sound. It takes text, first and last frames, or an omni-reference mode with up to nine images, three video clips, and three audio clips. ComfyUI workflow templates are linked from the model card.

Limits: The MiniMax H3 Community License excludes "the European Union, the United Kingdom, the Republic of Korea and the United States of America," and MiniMax runs a separate application form for those regions. Products earning more than $20 million a year need written permission. The repository is 498GB, MiniMax publishes no VRAM figure, and its example server runs on four GPUs. The 2K upscaling module hasn't been open-sourced.

Best for: Long clips built from many references, for readers outside the excluded regions.

6. MAGI-1: best for extending clips

Sand AI's MAGI-1 generates video chunk by chunk, 24 frames at a time, instead of all at once.

Strengths: Extension is built in. MAGI-1 does text-to-video, image-to-video, and video-to-video, continuing a clip from a prefix video, at arbitrary resolutions. It's Apache 2.0, ComfyUI custom nodes have existed since May 2025, and Sand AI released MAGI-1.1 24B weights in June 2026.

Limits: The 24B models want eight H100s, or eight RTX 4090s with fp8 quantization. The 4.5B model needs 24GB, or 12GB with its distilled, quantized build and a reduced window size. Output is video only.

Best for: Longer shots on a 24GB card, built up in pieces.

7. SkyReels V3: best for avatars and references

SkyReels V3, from Skywork, arrived in January 2026 as three task models, not a general text-to-video model.

Strengths: A 14B reference-to-video model puts subjects from reference images into a 720p clip, a 14B extension model continues a clip by 5 to 30 seconds, and a 19B talking-avatar model animates a portrait from audio up to 200 seconds long. For cards under 24GB, the repo supports FP8 weights with block offloading and lower resolutions.

Limits: You can't prompt it from text alone, and the repo doesn't mention ComfyUI. The Skywork Community License allows commercial use but asks you not to deploy the model in internet services without security review.

Best for: Talking heads and multi-subject clips built from photos.

Other open video models you'll still see

Older models remain on Hugging Face. Genmo's Mochi 1 is a 10B, Apache 2.0 model at 480p that needs about 60GB on one GPU, under 20GB in ComfyUI. CogVideoX1.5-5B makes 5- or 10-second clips at 1360 by 768 from about 10GB in BF16, under its own CogVideoX license. StepFun's Step-Video-T2V is 30B and MIT-licensed but peaks near 78GB. Open-Sora 2.0, an 11B Apache 2.0 model from March 2025, peaked at 44.3GB to 60.3GB per GPU in its maker's H100 tests.

Open licenses vs. community licenses

Apache 2.0 and MIT let you use, modify, and sell work from a model without revenue limits or content rules, which covers Wan 2.2, Kandinsky 6.0, MAGI-1, Mochi 1, and Step-Video. Community licenses add conditions: a revenue cap (LTX-2.5 at $10 million, MiniMax H3 at $20 million), excluded regions (HunyuanVideo 1.5 and MiniMax H3), or an acceptable use policy that binds your output (LTX-2.5). If you plan to train add-on files on one of these bases, our LoRA explainer covers how the base model's license carries over.

Before you download

• Check your VRAM first: Our VRAM guide breaks down what each card size runs and how offloading trades memory for speed.

• Check your disk: Video setups run 18GB to 43GB each before outputs, and full repositories go far higher. See how much storage local AI models need.

• Pick a mode: Most of these models do both text-to-video and image-to-video. Our guide to the two modes covers when to start from a still.

• On a Mac: New video models usually ship with CUDA code, and Lightricks lists only NVIDIA GPUs. The Mac vs. PC guide explains what runs on Apple Silicon.

Open-weight video models FAQ

What is the best open-weight video model?

Wan 2.2 for most local setups, because it's Apache 2.0, runs on small cards in ComfyUI, and has native workflows for text-to-video, image-to-video, and first-and-last-frame. LTX-2.5 is the pick for audio and 4K if you have 32GB of VRAM and qualify under its license.

Which open video models generate audio?

LTX-2.5, Kandinsky 6.0, and MiniMax H3 generate sound with the picture. Kandinsky 6.0 is the only one of the three under an open license (MIT). Wan 2.2's speech-to-video model and SkyReels V3's avatar model take audio as input and animate to it.

Is Wan 2.5 open source?

No. Alibaba's Wan organization on Hugging Face lists Wan 2.2 models and add-ons, with no Wan 2.5 or later video weights as of November 2026. Those versions run on hosted services.

Can I run an open video model on 8GB of VRAM?

Yes, with limits. ComfyUI says Wan 2.2's 5B model fits 8GB with offloading. Expect long waits, and plenty of system RAM helps.

Can I use open video models commercially?

Wan 2.2, Kandinsky 6.0, and MAGI-1 allow it with no revenue cap or regional limit. LTX-2.5 is free under $10 million in annual revenue, MiniMax H3 needs permission above $20 million, and HunyuanVideo 1.5's license doesn't cover the EU, the UK, or South Korea.