The open video model I would start with is Wan 2.2: it is Apache-2.0, ComfyUI ships templates for it, and its 5B model makes a 5-second 720P clip on any 48 GB card. The differences I weigh first are licences and memory, because a quality ranking depends on your prompts. HunyuanVideo 1.5 needs only 14 GB with offloading but is not licensed in the EU, UK or South Korea, LTX-2.3 and LTX-2.5 add audio but require a paid licence from companies with $10M or more in annual revenue, and MiniMax H3's licence excludes the United States, where every QuantaCloud region is.
The open video models, compared#
The table lists the families with open weights and native ComfyUI templates in September 2026, with the files each ComfyUI template loads, summed from the Hugging Face byte sizes (our calculation).
| Model | Released | Parameters | Licence | Files the ComfyUI template loads | The maker's memory guidance |
|---|---|---|---|---|---|
| Wan 2.2 TI2V-5B | July 2025 | 5B | Apache-2.0 | 18.14 GB | 24 GB for 720P with offloading, 22.9 GB peak |
| Wan 2.2 T2V-A14B and I2V-A14B | July 2025 | 27B in two 14B experts, 14B active per step | Apache-2.0 | 35.58 GB with fp8 experts | 80 GB for 720P. Peaks of 41.3 GB at 480P and 59.8 GB at 720P |
| HunyuanVideo 1.5 | November 2025 | 8.3B | Tencent Hunyuan Community License | 29.00 GB for 720p text to video | 14 GB minimum with offloading |
| LTX-2.3, video with audio | March 2026 | 22B, Gemma 3 12B text encoder | LTX-2 Community License | 42.96 GB: fp8 checkpoint, text encoder, two LoRAs and an upscaler | No figure published |
| LTX-2.5, video with audio | August 2026 | 22B, Gemma 4 12B text encoder | LTX-2.x Community License | 44.91 GB with the prompt enhancer the template turns on, 39.71 GB without it | 32 GB of VRAM and 32 GB of RAM minimum, A100 80GB or H100 recommended |
| Kandinsky 5.0 Video Lite | September 2025 | 2B | MIT, with HunyuanVideo's VAE under Tencent's licence | 14.70 GB | No figure published |
| MiniMax H3, video with audio | August 2026 | 33B | MiniMax H3 Community License | Not listed here | Not licensed for use in the US |
Wan's figures come from its own reference code on one GPU, and Tencent's and Lightricks' from their own guidance. ComfyUI streams weights between system RAM and the GPU, so it runs most of these in less memory than the makers quote, and more slowly when it has to.
Older open models such as the first HunyuanVideo, LTX-Video 0.9 and Wan 2.1 still have ComfyUI templates. I would start new work on the newer versions above. The newest Wan releases, Wan 2.5 to 3.0, are not open weights: Wan-AI had published none of them on Hugging Face by 2026-09-28, and ComfyUI runs them only as paid API nodes.
What each licence allows#
The licence decides more than the benchmark, because the same clip can be fine for a hobby project and off-limits for your company.
| Licence | Commercial use | Where it applies | Other conditions |
|---|---|---|---|
| Apache-2.0 (Wan 2.2) | Yes | Everywhere | Wan claims no rights over generated content |
| MIT (Kandinsky 5.0 Video Lite's own weights) | Yes | Everywhere | Keep the copyright notice. The model encodes and decodes video with HunyuanVideo's VAE, which carries the Tencent licence in the next row |
| Tencent Hunyuan Community License (HunyuanVideo 1.5, and the HunyuanVideo VAE) | Yes, unless your products had over 100 million monthly active users at its release | Not the EU, the UK or South Korea | Outputs may not be used to improve other AI models |
| LTX-2 and LTX-2.x Community Licenses (LTX-2.3, LTX-2.5) | Yes below $10M annual revenue, a paid licence above it | No territory limit beyond sanctions law | Not in a product that directly competes with Lightricks' offerings without a separate licence. LTX-2.3's text encoder is built on Google's Gemma 3, which comes with Google's Gemma terms |
| MiniMax H3 Community License | Only inside its territory, with authorization above $20M yearly revenue | Not the EU, the UK, South Korea or the United States | Outputs may not be used to improve other AI models |
The MiniMax row is why H3 has no file sizes above. Its licence does not cover use in the United States, and QuantaCloud's regions are all in the US, so I leave it out. The Kandinsky row is a reminder that a workflow's licence is the sum of its files: ComfyUI's Kandinsky template loads its VAE from a HunyuanVideo repository, so Tencent's territory and user limits come with it. Read each licence on its model card before you build a product on a model, and treat this table as a summary, not legal advice.
Which GPU runs which model#
The rule I follow is to compare the template's files with the card's memory first, then the maker's guidance. Every QuantaCloud GPU has at least 48 GB, which is past every published minimum in the table except Wan's 80 GB guidance for A14B at 720P.
| GPU | Memory | What I would run on it |
|---|---|---|
| RTX A6000, RTX 6000 Ada, L40, L40S | 48 GB | Wan 2.2 5B at 720P, Wan 2.2 A14B at 480P, HunyuanVideo 1.5 and Kandinsky 5.0 Video Lite. LTX-2.5 meets its 32 GB minimum here, though its template's files come to 44.91 GB, so ComfyUI may offload some of them |
| A100 80GB, H100 PCIe | 80 GB | Wan 2.2 A14B at 720P, which Wan says needs at least 80 GB, and LTX-2.5 on the class of card Lightricks recommends |
| RTX PRO 6000 Blackwell | 96 GB | All of the above with room to spare, including A14B at 720P and LTX-2.5 with prompt enhancement |
| H200 NVL | 141 GB | Wan 2.2 A14B with its 28.58 GB fp16 experts instead of fp8, and longer clips |
Two details change the choice between cards of the same size. The Ada, Hopper and Blackwell cards can compute in FP8, while the Ampere RTX A6000 and A100 cannot, so ComfyUI computes fp8 models in 16-bit there, which is slower. And offloaded weights wait in system RAM: RTX A6000 1x offers had 24 to 64 GB of RAM on 2026-09-27, against 144 GB on the RTX PRO 6000 1x. Lightricks asks for at least 32 GB of RAM for LTX-2.5, so check the RAM column before you launch an RTX A6000 for it.
| GPU | Memory | From | Available now |
|---|---|---|---|
| RTX A6000 | 48 GB | $0.48/GPU-hr | Yes |
| RTX 6000 Ada | 48 GB | $0.78/GPU-hr | Yes |
| L40S | 48 GB | $1.09/GPU-hr | Yes |
| A100 SXM4 80GB | 80 GB | $1.49/GPU-hr | Yes |
| H100 PCIe | 80 GB | $2.59/GPU-hr | Yes |
| RTX PRO 6000 Blackwell | 96 GB | $2.39/GPU-hr | Yes |
| H200 NVL | - | Not listed | No |
The RTX PRO 6000 Blackwell page has its configurations, ComfyUI GPU requirements covers the image models, and how much VRAM you need covers the arithmetic for language models too.
ComfyUI support and the template's version#
Every model in the comparison table has a native ComfyUI template, so no custom nodes are needed: open the Templates sidebar and search for the model's name. The catch is the ComfyUI version. LTX-2.5 and MiniMax H3 arrived in August 2026, so they need a ComfyUI release from then or later, and the version inside QuantaCloud's ComfyUI template is not published yet. Check it before you download 45 GB of files:
curl -s http://127.0.0.1:8188/system_stats | python3 -c 'import json, sys; print(json.load(sys.stdin)["system"]["comfyui_version"])'
If it is older than the model, install a current ComfyUI yourself on the Bare Metal template, as running ComfyUI on a cloud GPU shows. LTX-2.5's repository is also gated: accept its terms on Hugging Face and set HF_TOKEN before you download.
How to choose#
The honest answer depends on three questions, in this order: may you use the model where you are, may you use it for your business, and does it fit the GPU.
For a commercial product built in the US, Wan 2.2 is the clean choice, because Apache-2.0 puts no revenue or territory limits on you. Kandinsky 5.0 Video Lite's own weights are MIT, but its VAE brings Tencent's terms with it, the same as HunyuanVideo's. Wan 2.2 A14B is the one I would pick for image-to-video, since it has a dedicated I2V model, and the 5B for fast drafts. Wan 2.2 in ComfyUI has the files, settings and GPU choices in detail.
For clips with sound, LTX-2.3 and LTX-2.5 generate audio with the video, and their licence is free until your company reaches $10M in annual revenue. HunyuanVideo 1.5, at 8.3B parameters with a 14 GB minimum, sits between Wan's 5B and the 22B LTX models, but its licence stops at the EU, UK and South Korean borders, which matters if your users are there.
Whatever you pick, each QuantaCloud instance starts with an empty disk, and stopping it deletes the disk with every model and clip on it. Keep a list of model URLs, download on launch, and copy clips off before you stop. Loading models on a new ComfyUI instance has a script for the downloads, and moving files to and from a GPU server covers getting clips off.
Questions about open video models#
What is the best open-source video generation model?
There is no single answer that holds for every prompt, so I choose on licence and memory. For most people that is Wan 2.2: Apache-2.0, native ComfyUI templates, a fast 5B model and a larger 27B mixture-of-experts pair.
Can I use these models commercially?
Wan 2.2, yes. HunyuanVideo 1.5 yes, outside the EU, UK and South Korea, unless your products had over 100 million monthly users at its release, and Kandinsky 5.0 Video Lite on the same Tencent terms, because it uses HunyuanVideo's VAE. LTX-2.3 and LTX-2.5 yes below $10M in annual revenue. MiniMax H3 is not licensed for use in the US at all.
What is the best GPU for AI video generation?
For the models above, a 48 GB card covers Wan 2.2 5B, HunyuanVideo 1.5 and Kandinsky 5.0 Video Lite. The 96 GB RTX PRO 6000 Blackwell is the one I would rent for Wan 2.2 A14B at 720P and for LTX-2.5, because Wan's own code peaks at 59.8 GB at 720P and Lightricks recommends an 80 GB card for LTX-2.5.
How much VRAM does HunyuanVideo 1.5 need?
Tencent gives 14 GB as the minimum with model offloading. ComfyUI's 720p text-to-video template loads 29.00 GB of files, less than the 48 GB of the smallest QuantaCloud GPU.
Are Wan 2.5 and Wan 3.0 open source?
No. They are available only through paid APIs, including ComfyUI's API nodes, and Wan-AI has published no weights for them.
My decision rule: start with Wan 2.2 on a 48 GB RTX A6000, move to the RTX PRO 6000 Blackwell when you need A14B at 720P or LTX's audio, and read the licence before you pick anything else. The ComfyUI page lists every GPU with the template and its live price.
Launch ComfyUI on an RTX A6000