7 Best Open-Weight AI Video Generation Models in 2026

A single photograph can become a moving scene, a written prompt can direct a camera shot, and newer video models can generate speech and environmental sound alongside the picture. Open-weight AI video models make these creative capabilities available through downloadable checkpoints that you can run in your own projects.

In this post, you’ll find the 7 best open video generation models for everything from product animations and short character performances to connected scenes and longer videos. The list includes text-to-video and image-to-video models, newer generators with synchronized sound, and specialized options for extending footage or developing audio-video productions.

Quick Overview

Model	Main Use
MiniMax H3	Reference-guided video with dialogue and sound
Wan 2.2	Text and image animation, character movement
HunyuanVideo 1.5	Local text-to-video and image-to-video
LTX-2.5	Multi-shot creation, audio and clip editing
Kandinsky 6.0 Video	Short videos with synchronized speech
LongCat-Video	Video continuation and longer scenes
MAGI-2 Preview	High-resolution audio-video production

How I Picked

I selected models with downloadable video-generation weights and public inference code. I checked the released capabilities and used independent evaluations from Arena and Artificial Analysis to assess candidates.

1. MiniMax H3

MiniMax H3 generates video and synchronized stereo sound from text, images, and multimodal references. Describe a character’s actions and dialogue in a prompt, animate a portrait, or use existing footage to direct a new performance. The reference-guided variant accepts multiple images, video clips, and audio, so a sequence can follow the appearance and movement established in the supplied material.

The first-and-last-frame variant gives you control over both ends of a shot. Supply an opening image to animate its composition, or provide a second image to specify where the movement should finish. For character scenes, the reference-guided variant can draw on multiple views of a person and an audio performance when generating the new picture and soundtrack.

The downloadable H3-Base checkpoints generate 768p clips at 24 fps, with durations from 4 to 15 seconds. Diffusers and SGLang provide local inference options.

Model access: Open H3-Base weights.

License: MiniMax H3 Community License, with authorization required for deployment in the US, EU, UK, and South Korea and for certain high-revenue commercial services.

Official repository · Models and downloads

2. Wan 2.2

Wan 2.2 generates video from text and animates still images. A text prompt can specify the subject, setting, lighting, camera movement, and action. In image-to-video mode, a photograph or illustration establishes the opening composition, and the model generates motion from that frame.

The TI2V-5B checkpoint handles both tasks in one model and produces 720p-class footage at 24 fps. The larger T2V-A14B and I2V-A14B checkpoints are dedicated to text-generated footage and image animation, respectively. You can run the main models through ComfyUI or Diffusers, and model-specific LoRA workflows are available for adapting their output to particular subjects and styles.

Wan also has dedicated models for character animation. S2V-14B takes a character image and an audio recording to animate a speaking or performing subject. Wan2.2-Animate-14B uses a driving video to transfer movement and facial expressions to another character or replace the performer in a clip.

Local setup: The reference TI2V-5B configuration uses 24GB VRAM with offloading.

License: Apache 2.0 for the main video-generation code and weights.

Official repository · TI2V-5B weights · T2V-A14B · I2V-A14B

3. HunyuanVideo 1.5

HunyuanVideo 1.5 generates video from written prompts and starting images using an 8.3B-parameter model. Its 480p and 720p checkpoints can create an original scene from text or animate a photograph, illustration, or character design.

For image animation, the 480p step-distilled checkpoint uses 8 or 12 sampling steps. You can generate several versions of a subject’s movement or camera motion and compare the resulting clips. The standard 720p generation options and an optional super-resolution stage provide higher-resolution output, including 1080p after upscaling.

You can run HunyuanVideo 1.5 through Python, Diffusers, or ComfyUI. Training scripts and LoRA adapters provide ways to adapt the model to a recurring character or a particular illustration style.

Local setup: Linux and NVIDIA CUDA, starting at about 14GB VRAM with offloading.

License: Tencent Hunyuan Community License, which excludes the EU, UK, and South Korea.

Official repository · Model downloads

4. LTX-2.5

LTX-2.5 generates synchronized video and audio, including dialogue, sound effects, and sequences with multiple shots. A text prompt can describe several moments of action, or an image can establish the opening look. Multi-shot controls guide changes in camera position and composition while the model generates picture and sound together.

Keyframe conditioning lets you specify important moments within a sequence. LTX-2.5 also accepts an audio recording as input for video generation, which can be used to produce moving footage based on speech or another soundtrack.

For revisions, Retake regenerates a selected time range in an existing clip from a new prompt. IC-LoRA pipelines handle image-guided and video-to-video transformations. Diffusion Fidelity Rendering generates additional keyframes and detail passes, with spatial and temporal upscaling for higher-resolution output.

LTX-2.5 has 22B development and distilled transformers. The development model has training and fine-tuning workflows, including LoRA. The distilled version uses fewer denoising steps for inference. Both can be used in Python pipelines, and ComfyUI provides a node-based generation and editing workflow.

License: LTX-2.x Community License, including commercial use for eligible businesses under $10 million in annual revenue.

LTX-2.5 overview · Official repository · Model downloads

5. Kandinsky 6.0 Video

Kandinsky 6.0 Video creates five-second clips with synchronized audio from a text prompt or a starting image. It can generate a character speaking, a short exchange between characters, or an action accompanied by environmental sound. Speech and lip movements are generated together with the video.

The family includes Kandinsky 6.0 Lite at 3B parameters and Kandinsky 6.0 Pro at 29B. Both produce synchronized 44 kHz audio. Text conditioning builds the scene from a description, while image conditioning animates an existing portrait, illustration, or composition.

Distilled checkpoints provide an inference option with reduced sampling steps. You can run the released variants through Diffusers, ComfyUI, or SGLang. An additional super-resolution model can upscale the resulting clip to 1920 × 1080.

License: MIT

Official repository · Lite distilled weights · Pro distilled weights

6. LongCat-Video

LongCat-Video generates new footage and continues an existing clip beyond its original ending. After creating the opening scene from a text prompt or still image, you can feed the preceding footage into the continuation model to generate the next segment. A tracking shot can follow a subject farther, or an action can develop over several connected segments.

The 13.6B base model handles text-to-video, image-to-video, and video continuation at 720p and 30 fps. Its continuation training uses preceding video to guide the movement and appearance of subsequent footage. Multiple extensions can build a longer sequence from the initial clip.

The LongCat-Video-Avatar 1.5 release handles audio-driven character animation. It takes a recording and character references to generate speaking or performing characters.

License: MIT

Official repository · Download base weights · Avatar 1.5 model

7. MAGI-2 Preview

MAGI-2 Preview generates ten-second video clips with synchronized sound from a text description or an opening image. The output includes the moving scene and its soundtrack in a single generation process, including dialogue or environmental audio specified by the prompt.

Its two-stage pipeline generates an initial video and refines the picture to 1080p in a second stage. MAGI-2 uses a mixture-of-experts model with 114 billion parameters in total and approximately 6 billion active parameters per token.

MAGI-2 Preview runs on a multi-GPU server, with a reference configuration using eight NVIDIA Hopper GPUs. A Docker image is available for deploying the inference code.

License: Apache 2.0 for MAGI-2 code and weights, with additional conditions for its Stable Audio Open component.

Official repository · Download weights

Which One Should You Choose?

For text-to-video and image animation on a workstation, Wan 2.2 has a 5B checkpoint that handles both tasks, with ComfyUI and Diffusers integrations. HunyuanVideo 1.5 has a reference offloading configuration starting at about 14GB of VRAM, plus a step-distilled image-to-video model and LoRA training tools.

For multi-shot video with synchronized sound and the ability to revise part of a clip, choose LTX-2.5. Kandinsky 6.0 Video generates five-second clips with speech and lip synchronization. MiniMax H3 accepts first-and-last-frame controls and multimodal references for character and dialogue scenes.

For extending a scene through successive video segments, use LongCat-Video. For ten-second audio-video clips refined to 1080p on an eight-GPU server, MAGI-2 Preview provides the inference pipeline.

Frequently Asked Questions

Can I run an open-weight AI video model on an 8GB or 12GB GPU?

A few community workflows use quantized weights, offloading, and reduced output settings to run smaller or compressed video models on GPUs in that range. The normal full-precision reference configurations are often much larger. If you have 12GB, focus on a model-specific quantized ComfyUI setup. HunyuanVideo 1.5 has a reference configuration starting at about 14GB with offloading. Wan 2.2 TI2V-5B uses 24GB with offloading in its reference configuration.

Do I need ComfyUI to generate videos locally?

No. ComfyUI provides a graphical, node-based way to connect a prompt, model, reference images, and video export. Models such as Wan 2.2, HunyuanVideo 1.5, and LTX-2.5 also have Python-based inference options. Pick ComfyUI if you prefer arranging and reusing generation steps visually. Use Python when you want to automate batches or integrate generation into an application.

Which open-weight models generate video with sound?

MiniMax H3, LTX-2.5, Kandinsky 6.0 Video, and MAGI-2 Preview generate synchronized picture and sound. Their creative controls differ: H3 can use audio and other references, LTX-2.5 includes audio-driven and editing pipelines, and Kandinsky focuses on short clips with synchronized speech. Wan 2.2’s standard text-to-video and image-to-video checkpoints generate silent footage. Its S2V variant animates visuals to audio you supply.

Can I turn a photograph or illustration into a video?

Yes. Image-to-video generation uses your uploaded still image as a starting point and generates movement around it. Wan 2.2 and HunyuanVideo 1.5 both animate an input image. MiniMax H3 accepts first and last frames to control the transition, and LTX-2.5 uses images and keyframes to guide an audio-video sequence.

How can I make AI-generated videos longer than a few seconds?

Generate consecutive segments that inherit information from the preceding footage. LongCat-Video has a dedicated video-continuation model for this task and can build minutes-long sequences through repeated extensions. Other models can create individual clips that you edit together, but joining independent clips usually requires extra attention to character appearance, camera movement, and continuity.

Are open-weight AI video generators free to use?

The model weights are available to download with no per-video subscription charge. Generating footage still consumes GPU resources. On your own hardware, the main cost is your computer and electricity. On rented hardware, you pay for GPU time. Free access to weights also does not imply unrestricted commercial use, since the license attached to each model determines usage permissions.

Can I use AI videos from these models in commercial projects?

Commercial rights depend on the model and any third-party components in the production pipeline. Wan 2.2, Kandinsky 6.0 Video, and LongCat-Video release their main model weights under permissive licenses. LTX-2.5, MiniMax H3, and HunyuanVideo 1.5 use community terms with additional conditions. MAGI-2 includes an audio component with its own license. Commercial use of that component requires registration with Stability AI, and businesses above $1 million in annual revenue need an enterprise license. Check the linked license for the model you plan to use in client work or a commercial service.

Can I train a LoRA or customize a video model for a character or style?

Yes, for model families that have released compatible training tools. Wan 2.2 has community training and LoRA integrations, HunyuanVideo 1.5 releases training scripts, and LTX-2.5 supports LoRA and development-model fine-tuning. An adapter trained for one model or checkpoint must be used with the corresponding inference model. LoRA training can encode a recurring subject or illustration style for a series of clips.

Why are Seedance and FLUX 3 Video not included?

The selection is for video-generation models with publicly downloadable weights and inference code. Seedance and FLUX 3 Video are important video products, but their available video-generation services do not meet that model-download requirement for this collection. They belong in a general list of hosted AI video generators.