6 open-weight AI video models that make sound, ranked by GPU

Six open-weight AI video models can now generate their own synchronized sound, and this list sorts them by the hardware each one needs to run. The ranking runs on three sourced criteria, in order: the lowest minimum GPU that a maker or its official ComfyUI pack states, then how permissive the weights license is, then the highest native resolution the open release reaches. Almost every figure below comes from the maker's own repository, model card or paper, with a few attributed to named outlets, read on 13 October 2026.

The spread is wide. The top entry runs its small model on a consumer graphics card and the last one needs a data-center H800. Licenses split the same way, from MIT and Apache-2.0 at the open end to community licenses that charge above a revenue threshold. Ranked by the machine you actually have.

1. Kandinsky 6.0 Video

The Kandinsky lab released Kandinsky 6.0 Video on 4 October 2026 under the MIT license, as a family of 3B Lite and 29B Pro diffusion models. Both generate 5-second clips with synchronized 44 kHz audio, including lip-sync, from text or image prompts, and a plug-in super-resolution model raises the output to Full HD 1920x1080, per the technical report.

Kandinsky leads because it is the only entry whose maker documents the small model on consumer cards. The repo's own performance table lists working times for a 5-second clip on the RTX 4090, RTX 5060 Ti, RTX 5080, RTX 5090 and RTX PRO 6000, driven through block offload, plus the A100 80 GB and H100 through module offload. Lite on an RTX 4090 takes 422 seconds at HD and 578 seconds at Full HD; the same Lite model on an RTX 5060 Ti takes 1328 and 1774 seconds.

The architecture is a dual-stream CrossDiT that joins a pretrained video stream to a newly trained audio stream through bidirectional cross-attention, which the report says aligns the two modalities in time and meaning. Kandinsky 6.0 arrived in Diffusers, ComfyUI, vLLM-Omni, SGLang and FastVideo on 6 October 2026, and there is a Hugging Face Space for the Pro Distill model.

• Maker: Kandinsky lab
• License: MIT
• Native audio: 44 kHz synchronized audio with lip-sync
• Max native resolution: Full HD 1920x1080 via a plug-in super-resolution model
• Minimum hardware: consumer GPUs, RTX 5060 Ti up to RTX 5090, via block offload
• Try it: https://github.com/kandinskylab/kandinsky-6

2. Ovi 1.1

Ovi 1.1 is next because its documented floor is 32 GB and it carries the most permissive license at that tier. Character AI's Ovi 1.1 generates 10-second or 5-second video at 960x960 and 24 fps from text or text plus image, with synchronized audio, and its weights are Apache-2.0.

The model card states that 32 GB of GPU memory is the minimum to run Ovi, and that fp8 or qint8 with CPU offload brings peak use down to about 24 GB with slight quality degradation. On the maker's own table, a single GPU with CPU offload and FlashAttention-3 enabled peaks near 32 GB and finishes a 121-frame 720x720 clip at 50 steps in about 140 seconds; that row assumes Ovi's optional image-generation model is switched on, and the card lists different memory and timings for other configurations.

Ovi's audio branch is a 5B model pretrained from scratch on in-house audio, and the video branch initializes from Wan2.2. There is a free demo Space, and a ComfyUI path now works through the WanVideoWrapper.

• Maker: Character AI and Yale University
• License: Apache-2.0
• Native audio: synchronized audio from a 5B audio branch, speech and sound effects
• Max native resolution: 960x960 at 24 fps
• Minimum hardware: 32 GB VRAM, about 24 GB with fp8 or qint8 plus CPU offload
• Try it: https://github.com/character-ai/Ovi

3. LTX-2.5

LTX-2.5 also documents a 32 GB floor, so the tiebreak on this list is the license: LTX-2.5 sits under the LTX-2.x Community License, free for commercial and production use only under $10M in annual revenue, where Ovi's weights are plain Apache-2.0. Above that threshold Lightricks sells a paid commercial agreement.

Lightricks' ComfyUI documentation asks for a CUDA-compatible GPU with 32GB or more VRAM and 100GB of free disk space, and the LTX blog lists the same 32 GB GPU floor alongside 32 GB of system RAM and CUDA 13.2 or higher. Lower-VRAM INT8 ConvRot files, or an NVFP4 distilled transformer on Blackwell cards, cut the memory need.

LTX-2.5 generates synchronized video and audio from text, image and video inputs, with separate VAEs for the picture and the soundtrack. The official ComfyUI templates run a two-stage pipeline with a 2x spatial upscaler for Text-to-Video and Image-to-Video, while Lightricks documents First-Frame/Last-Frame as a single-stage path without the upscaler. Creative AI Show puts the ceiling for that upscaler pipeline at 4K.

• Maker: Lightricks
• License: LTX-2.x Community License, free under $10M annual revenue
• Native audio: synchronized soundtrack decoded by a dedicated audio VAE
• Max native resolution: raised by a 2x spatial upscaler in the two-stage pipeline (Creative AI Show lists 4K)
• Minimum hardware: 32GB+ VRAM, 100GB disk and CUDA 13.2+, with INT8 ConvRot or NVFP4 files lowering it
• Try it: https://github.com/Lightricks/LTX-2

4. MiniMax H3

MiniMax H3 is a 33B dense transformer that predicts video and audio latents in one forward pass, MiniMax publishes no minimum-VRAM figure for it, so the tiebreak falls to its tighter license, and that is why it ranks fourth. The model card lists 4-15 second output, 24 fps, 32 kHz stereo audio and resolutions up to 2K under the MiniMax H3 Community License.

The open weights generate at a 768-pixel short edge, 1344x768 at 16:9 and about one megapixel, and Comfy's H3 guide says 2K output comes from MiniMax's hosted service or a separate upscale pass. MiniMax's card notes that the roughly 13B-parameter AdaLN branch can be precomputed and cached and does not need to be loaded for inference-only deployment, which is what the pruned builds exploit.

Comfy-Org ships pruned INT8 ConvRot and FP8 checkpoints for ComfyUI, alongside the full bf16 weights. Comfy's guide states that commercial use of locally generated outputs requires a MiniMax commercial license, available through Comfy as the official reseller.

• Maker: MiniMax
• License: MiniMax H3 Community License, with local commercial use needing a paid license via Comfy
• Native audio: 32 kHz stereo generated with the video in one pass
• Max native resolution: 768-pixel short edge from the open weights; 2K runs on MiniMax's hosted service
• Minimum hardware: 33B dense model; MiniMax publishes no minimum VRAM, pruned INT8 ConvRot builds come from Comfy-Org, and the ~13B AdaLN branch can be cached at inference
• Try it: https://huggingface.co/Comfy-Org/MiniMax-H3

5. Prism

Prism jumps the floor to data-center GPUs, but it is the only entry here that trains and generates joint video-audio natively at 2K, which places it above the 80 GB singles. From Fudan University, Tencent Hunyuan and Zhejiang University, Prism is an image-to-video model that generates synchronized speech, sound effects and music in the same pass, at native 720p, 1080p and 2K, with MIT-licensed weights on top of an Apache-2.0 MOVA-360p base.

The hardware is the catch. Prism's own project page and code repo state that 720p inference runs on a single NVIDIA 80 GB GPU, while 1080p and 2K each need four or more 80 GB GPUs, and training at those resolutions needs 64 or more. Creative AI Show reports about 143 GB of download for the stable checkpoint plus its MOVA base.

The method behind it is the paper "Dynamic Sparse Attention for Native 2K Joint Video-Audio Generation Model Training" (arXiv:2610.05416, 4 October 2026), which reports a 2.5x training speedup over full attention. Two preview variants are out, preview-alpha for stability and preview-beta for motion; there is no ComfyUI node yet.

• Maker: Fudan University, Tencent Hunyuan and Zhejiang University
• License: MIT weights on an Apache-2.0 MOVA-360p base
• Native audio: joint speech, sound effects and music generated with the picture
• Max native resolution: 2K native, with 720p and 1080p modes
• Minimum hardware: one 80 GB GPU at 720p, four or more 80 GB GPUs at 1080p and 2K
• Try it: https://github.com/Tencent-Hunyuan/Prism

6. Helix World

Helix World is the only entry that runs faster than real time, but it needs a data-center card and outputs at a much lower resolution than Prism, which lands it sixth. Noiz AI and the Hong Kong University of Science and Technology built it as a real-time interactive audio-visual world model that runs on a single NVIDIA H800 at 24 frames per second, turning an image plus keyboard input into an explorable scene with camera-aligned stereo.

The maker's blog post reports a steady-state real-time factor of 0.77 at 768x512, including video and audio decoding, and a spatial-acoustic consistency score of 41.8 on its HelixBench against 12.7 for EchoWM. The code repo recommends 80 GB of VRAM and says a reviewed BF16 single-GPU run uses about 70 GB; the paper reports no consumer-GPU figures.

Helix World builds on the open-source LTX-2.3 audio-visual model, adds 6-DoF camera and keyboard control, and distills the result into a few-step causal student for streaming. The code is Apache-2.0 and the weight license ships with the preview checkpoint.

• Maker: Noiz AI and HKUST
• License: Apache-2.0 code, with the weight license published with the preview checkpoint
• Native audio: camera-aligned stereo generated with the scene
• Max native resolution: 768x512 at 24 fps
• Minimum hardware: one NVIDIA H800, 80 GB recommended and about 70 GB BF16
• Try it: https://github.com/NoizAI/HelixWorld