Video generation is the workload where the model card lies the most, and it lies in the same direction every time. The card gives you the parameter count of the diffusion model, the commercial-use page gives you a clean Apache 2.0, and the inference script then asks for a text encoder, a 3D VAE, and a temporal attention stack that the parameter count leaves out. This guide gives the documented memory figures, and it puts licensing first, because one well-known open video model excludes the EU, UK and South Korea, and another caps free commercial use at a monthly visit threshold.
The mistake people bring from image gen
A text-to-video model is not one file. It is a pipeline: a diffusion transformer, an LLM text encoder, a 3D VAE for video, and a temporal module that connects frames. The text encoder and the 3D VAE are the parts the parameter count leaves out.
Wan2.2-TI2V-5B at 5B parameters sounds like an easy fit on a 24 GB card. The published
model card runs its 24 GB example with
--offload_model True --convert_model_dtype --t5_cpu, and says those options can be removed
only on a GPU with at least 80 GB. HunyuanVideo at 13B parameters lists a 45 GB minimum at
544x960 and 60 GB at 720x1280, tested on a single 80 GB GPU, and its FP8 weights save about
10 GB.
The encoder offloads cheaply here in a way it does not for an LLM, because the encoder runs
once per prompt rather than once per token. Offloading the encoder to CPU after
conditioning, as Wan’s --t5_cpu flag does, brings peak memory closer to the diffusion
model alone than to the sum of all parts.
Licences first, because this is where the traps are
The strongest open video models fall into three camps: permissive Apache 2.0, conditional Open Weights with revenue or visit caps, and custom community licences with territory or monthly-active-user clauses.
| Model | Params | Licence | Commercial use |
|---|---|---|---|
| Wan2.2-T2V-A14B | 27B total / 14B active | Apache 2.0 | Yes |
| Wan2.2-TI2V-5B | 5B | Apache 2.0 | Yes |
| LTX-Video 0.9.8 (13B dev / mix / distilled) | 13B | LTX-Video Open Weights 0.X | Yes, with a $10M revenue trigger |
| LTX-Video 0.9.8 (2B distilled) | 2B | LTX-Video Open Weights 0.X | Yes, with a $10M revenue trigger |
| LTX-Video 0.9.6 (2B distilled) | 2B | LTX-Video Open Weights 0.X | Yes, with a $10M revenue trigger |
| Mochi 1 preview | 10B + 362M VAE | Apache 2.0 | Yes |
| Allegro | 2.8B DiT + 175M VAE | Apache 2.0 | Yes |
| CogVideoX-5b | 5B (per name) | CogVideoX Community | Free after registration, under 1M visits/month |
| HunyuanVideo | 13B | tencent-hunyuan-community | Royalty-free, EX EU/UK/SK, MAU cap at 100M |
| Stable Video Diffusion img2vid | not stated on card | stable-video-diffusion-community | Conditional on Stability AUP |
Two patterns matter. CogVideoX-5b is free for commercial use only after registering for a basic commercial licence and only below 1 million visits per month, per the published LICENSE; above that the licence says to contact the licensor’s business team for more commercial licences. HunyuanVideo is royalty-free but its LICENSE excludes the EU, UK and South Korea entirely, and any product above 100M monthly active users has to ask Tencent for a separate licence granted at their discretion. LTX-Video is permissive for small operators but its Open Weights License requires a paid commercial licence once you cross $10M annual revenue, with double-fee liquidated damages for unauthorised commercial use as a large entity.
If the output is going anywhere near a commercial deliverable in the EU or UK, HunyuanVideo is out. The Apache 2.0 models (Wan 2.2 in both variants, Mochi 1 and Allegro) carry no territory, revenue or visit conditions.
By VRAM tier
The tiers below are based on each model’s own documentation at the model card’s native resolution and clip length. Several of the official figures already assume CPU offload or other memory savers, which cost real speed; where they do, that is said.
8 GB. This tier is barely usable for text-to-video at any meaningful resolution. The LTX-Video model card describes the 2B 0.9.8 distilled build as “Ideal for light VRAM usage” and the 2B 0.9.6 distilled build as “15× faster, real-time capable”, but publishes no memory figures. CogVideoX-5b is listed at “starting from 4.4GB” with INT8 (torchao) in diffusers per the model card, at 720x480 and 6 seconds at 8 fps, with every diffusers memory optimisation (including CPU offload) switched on; the card puts peak use at about three times higher without them. Treat both as the entry point for hobby work, not for production.
12 GB. LTX-Video 2B distilled is the family’s low-memory route; the card says the model works best under 720 x 1280 and below 257 frames but gives no VRAM floor, so the 13B builds on a 12 GB card are offload territory. Stable Video Diffusion generates 14 frames at 576x1024; its card times a clip at about 100 seconds on an A100 80GB and publishes no lower memory floor.
16 GB. CogVideoX-5b at BF16 is listed at “starting from 5GB” per the model card
with diffusers optimisations on; at roughly three times that peak without them, a 16 GB card
is the comfortable home for it. Allegro (a 2.8B video transformer) uses
9.3 GB in BF16 with CPU offloading and 27.5 GB
without, which is the cleanest 16 GB option when offload is acceptable. Wan2.2-TI2V-5B is
the right card-class target at 24 GB, not 16 GB: the Wan2.2-TI2V-5B card
documents 24 GB as the floor even with its --offload_model, --convert_model_dtype and
--t5_cpu flags on, and publishes no 16 GB configuration.
24 GB. This is the meaningful consumer tier. Wan2.2-TI2V-5B at 5B parameters runs on a GPU with at least 24 GB (the card’s example is an RTX 4090) at 720p and 24 fps with the offload flags turned on, and the card says it generates a 5-second 720P video “in under 9 minutes on a single consumer-grade GPU” per the model card. Mochi 1 at BF16 takes about 22 GB per the model card, or under 20 GB through the ComfyUI implementation. Wan2.2-T2V-A14B at 14B active is documented for a GPU with at least 80 GB, with offload flags suggested on out-of-memory errors and no smaller floor published.
32 GB and up. HunyuanVideo at 13B parameters needs a minimum of 60 GB at 720x1280 and 45 GB at 544x960 (129 frames), with 80 GB the recommended target, per the model card; the FP8 build saves about 10 GB. Wan2.2-T2V-A14B at 80 GB is the documented single-GPU figure. Running Wan2.2-TI2V-5B without its offload flags also takes an 80 GB card, per its card.
What to actually run
If the work is commercial and global: Wan2.2-TI2V-5B. It is Apache 2.0, ungated, runs on a single 24 GB consumer card with offload, and supports both text-to-video and image-to-video in one model. Only Allegro (2.8B, Apache 2.0) is smaller among the permissive entries, and this Allegro release is a text-to-video model producing 6 seconds at 15 fps.
If the target is a 24 GB card and a second Apache 2.0 option: Mochi 1. Apache 2.0, ungated, 22 GB at BF16 in diffusers per its card, “less than 20GB” through ComfyUI, and 480p native output.
If you need the lowest documented memory floor: CogVideoX-5b from 5 GB BF16 with diffusers optimisations on, provided your product registers for the basic commercial licence and stays under 1 million visits per month. Above that cap, the licence requires a further commercial licence.
If you are in the EU or UK, do not use HunyuanVideo under any workflow. Its published LICENSE names the EU, UK and South Korea as excluded territory and authorises use outside that set only.
One thing this guide does not do is rank these models by output quality. No primary benchmark was verified for this update, and video quality comparisons are even more prompt-dependent and temporally-coherent than image ones. The ordering above is by what fits, what you are licensed to do, and what architecture the model is built for. Treat any confident quality ranking you read elsewhere as weaker evidence than rendering twenty clips from your own prompts and watching the motion.
For the rest of the local stack, the VRAM index covers chat, coding, vision and image, the image-gen VRAM guide explains the same text-encoder trap for diffusion image models, and which quantization to use covers the weight formats behind builds such as the FP8 weights above.