Video generation GPU requirements: Wan 2.2, HunyuanVideo and LTX on your own GPU servers

Video generation GPU requirements in short

The GPU requirements for AI video generation follow from the model, the resolution and the clip length. The large open models generate 720p clips on one card with 80 to 96 GB: Wan 2.2’s text-to-video model “can run on a GPU with at least 80GB VRAM”, its repository says, which a 96 GB RTX PRO 6000 or a 141 GB H200 NVL provides. Smaller models and offloading to system memory bring the figure down to 24 GB for Wan2.2-TI2V-5B and 14 GB for HunyuanVideo 1.5.

Video models are diffusion transformers, like the image models in our FLUX GPU requirements guide, but each denoising step works on a whole sequence of frames. MLPerf® Inference v6.1 includes a text-to-video benchmark with Wan2.2-T2V-A14B, and MLCommons’ task force describes its clips as 81 frames at 720×1280 and 16 frames per second, 5 seconds long, generated with 20 denoising steps. Memory decides which model and resolution fit on a card, and compute decides how long a clip takes.

For a company, the licence is the second filter after memory. As read on 16 October 2026, Wan 2.2 is under Apache 2.0, the HunyuanVideo 1.5 and MiniMax H3 licences exclude the European Union, and the LTX-2.x licence asks companies above a revenue threshold for a paid licence.

Open video models in October 2026: parameters, memory and licence

Wan 2.2 was released with weights and inference code on 28 July 2025. Its two large models are mixture-of-experts designs: “Each expert model has about 14B parameters, resulting in a total of 27B parameters but only 14B active parameters per step”, with a high-noise expert for the layout and a low-noise expert for the details. We found no published weights for a later Wan version. As of 16 October 2026, the most recently updated repositories in the Wan-AI organisation on Hugging Face are the Wan2.2-Animate-2-14B models for character animation from a driving video, released on 7 August 2026.

Tencent released HunyuanVideo 1.5, an 8.3-billion-parameter diffusion transformer, on 20 November 2025; it generates 480p or 720p with 121 frames by default and upscales to 1080p separately; we found no later HunyuanVideo weights. Lightricks names LTX-2.5 as the recommended model of its LTX-2 repository. Its checkpoint files carry a 22B transformer, it generates “synchronized, high-fidelity video and audio”, and its text encoder is Gemma 4 12B fine-tuned for LTX, as the repository describes it. MiniMax H3, under a licence dated 2 August 2026, is a 33B dense transformer that generates video with stereo audio for 4 to 15 seconds and up to 2K.

GitHub repositories and Hugging Face model cards of Wan-Video, Tencent-Hunyuan, Lightricks and MiniMaxAI, and their licence files, read on 16 October 2026.

Licences for company use: Apache 2.0, revenue thresholds and EU exclusions

Wan’s repository states: “The models in this repository are licensed under the Apache 2.0 License.” Its licence file is the unmodified Apache License 2.0, which grants commercial use with no territory or revenue condition.

The HunyuanVideo 1.5 licence defines its Territory as “the worldwide territory, excluding the territory of the European Union, United Kingdom and South Korea”, and states in capitals that it “DOES NOT APPLY IN THE EUROPEAN UNION, UNITED KINGDOM AND SOUTH KOREA”. Above 100 million monthly active users, a licensee must request a licence from Tencent. Outputs may not be used “to improve any other AI model” other than Tencent Hunyuan and its derivatives.

The MiniMax H3 Community License lists “the European Union, the United Kingdom, the Republic of Korea and the United States of America” as Excluded Territories and invites interested parties there to contact MiniMax about a licence. It adds: “Any such use outside the Applicable Territory is not authorized by this Agreement.”

The LTX-2.x Community License, dated 11 August 2026, covers every LTX-2.5 version released since that day. Companies whose annual revenue reaches a threshold set in Section 2.1 count as Commercial Entities, with “all subsidiaries, affiliates, and companies under common Control” considered collectively, and “are required to obtain a paid license for any use” beyond the non-commercial purposes of Section 2.2. Those purposes include testing and evaluation by a Commercial Entity “in a non-production or development environment”. Lightricks “claims no rights in the Output”, while models trained on generated outputs count as derivatives under the licence.

The use terms bind the clips as well. Wan’s repository says that use “must not involve sharing any content that violates applicable laws”, and the acceptable use policy attached to the LTX-2.x licence prohibits placing generated content in any context “without expressly and intelligibly disclaiming that the information and/or content is machine generated”.

Our comparison of open LLM licences for company use covers a territorial clause of the same kind in Meta’s Llama 4 use policy, which withholds rights to its multimodal models from companies in the EU. Whether a given use is covered by one of these licences is a legal assessment for the company’s legal department.

Wan 2.2 VRAM and hardware requirements

The single-GPU command in Wan’s repository generates a 1280×720 clip with --offload_model True and --convert_model_dtype, and the repository states that it runs on a GPU with at least 80 GB. If memory runs out, it suggests --t5_cpu as well, which moves the text encoder to the processor. The 72 GB RTX PRO 5000 is below it, and Wan publishes no figure for that case, so test the A14B models on one card with the three options before you size several.

Wan2.2-TI2V-5B uses a VAE with a compression ratio of 16×16×4 and generates 720p, which its card defines as 1280×704, at 24 frames per second. It runs on “a GPU with at least 24GB VRAM (e.g, RTX 4090 GPU)” with the offloading options, and on a card with at least 80 GB the card says you can remove them “to speed up execution”.

Offloading moves weights into system memory, so the server needs RAM for every model it offloads. Our RTX PRO 6000 vs H200 NVL comparison sets out memory, bandwidth and NVLink of the two cards.

HunyuanVideo VRAM and LTX video GPU memory

HunyuanVideo 1.5 gives a minimum of 14 GB, “measured with model offloading enabled”, and advises disabling offloading on cards with enough memory for faster inference. Its repository added a step-distilled 480p image-to-video model, run in 8 or 12 steps, and FP8 GEMM inference in December 2025. A sparse attention option “requires H-series GPUs”, which includes the H200 NVL but not the RTX PRO cards.

Lightricks gives no VRAM figure for LTX-2.5. By our arithmetic, the 22B transformer takes about 44 GB in BF16 and the 12B text encoder about 24 GB, which fits a 96 GB card before working memory for the frames. The repository offers --quantization fp8-cast, which converts BF16 checkpoints to FP8 on load and halves the transformer to about 22 GB, and --offload cpu for smaller cards. Its NVFP4 checkpoint, a distilled transformer, runs with nvfp4-prequant on Blackwell, which covers the RTX PRO cards but not the H200 NVL. For FP8 checkpoints on Hopper and later GPUs, the repository names fp8-scaled-mm.

Generation time per clip: published figures

HunyuanVideo 1.5 states that with the step-distilled 480p image-to-video model “a single RTX 4090 can generate videos within 75 seconds”. NVIDIA’s page lists Stable Video Diffusion, an older image-to-video model, at 2.82 videos per minute on one RTX PRO 6000 Server Edition and 1.34 on an L40S, without resolution or frame count. We found no time per clip for Wan 2.2, HunyuanVideo 1.5 or LTX-2.5 on an RTX PRO card from the model publishers or NVIDIA. Times for one Spark are in our article on DGX Spark beyond LLMs.

Multi-GPU inference: sequence parallel or one clip per card

Wan’s repository runs one clip across eight GPUs “using FSDP + DeepSpeed Ulysses”, with --ulysses_size 8 in its 720p example. HunyuanVideo 1.5 sets the GPU count with --nproc_per_node, eight in its example. The MiniMax H3 card shows SGLang with four GPUs and a Ulysses degree of 4. xDiT describes itself as “an inference engine designed for the parallel deployment of DiTs on a large scale”, supports Wan2.2, HunyuanVideo-1.5, LTX-2.5 and MiniMax-H3, and combines Ulysses and Ring attention into what it calls Unified Sequence Parallel.

Sequence parallelism shortens the wait for one clip, while a team that queues many clips gets more per hour when each card runs its own job and the cards exchange no data. Each card then holds a full copy of the model, as each card holds the full scene in a GPU render server, so the per-card memory figures above apply. The RTX PRO 6000 has no NVLink, so sequence parallel traffic between its cards runs over PCIe. The H200 NVL bridges 2 or 4 cards over NVLink, and an eight-card H200 NVL server forms two NVLink domains. The repositories publish no scaling figures for PCIe-only cards, so measure the gain before you size for it.

For the eight-GPU examples over NVLink in one system, NVIDIA’s DGX B300 page lists “8x NVIDIA Blackwell Ultra SXM” with 2.1 TB of GPU memory; our DGX B300 comparison sets it against an eight-card RTX PRO 6000 server.

For a video platform on eight NVLink-connected GPUs we supply the DGX B300 per project; the distribution first asks for the company, the project and the intended use. Name those three and your clips a day, and we check price and availability.

AI video generation server for a team

A marketing, training or product team sends clips to a queue, and ComfyUI, covered in our FLUX guide, queues jobs per instance. Size the server in four steps:

1. Choose the models the licence allows for your company and the resolution and clip length you need.
2. Take the memory from the model table above and pick the card that holds the model without offloading.
3. Measure the minutes per clip on one card at your settings, and divide the clips in the busiest hour by the clips one card delivers in an hour.
4. Check the wait: with one job per card, the number of people submitting at the same time often sets the card count before the daily volume does.

Estimates by Eurokommerz from the model table above and the repositories’ multi-GPU examples; card and DGX B300 memory from NVIDIA’s product pages, read on 16 October 2026.

Plan the model store as well. LTX-2.5’s downloads alone total about 66 GiB, and by our estimate a server that keeps three or four video models ready holds several hundred gigabytes of checkpoints on NVMe. Store the prompt, seed and model revision with every clip, so that a result can be reproduced and its licence traced.

We build servers with 4 or 8 RTX PRO 6000 Server Edition or H200 NVL cards. Describe your video workflows in the form below, with the clips a day and the number of people generating at peak.

What we supply

We supply the cards this article discusses, the RTX PRO 6000 in all three editions, the RTX PRO 5000 with 48 or 72 GB, the H200 NVL and the L40S, with manufacturer warranty on one EU contract and invoice; the GPU page lists each card’s memory and power. We build AI servers to order for video generation, assembled and burn-in tested, and we check the rack, power and airflow before we quote. We supply the DGX B300 per project, and approvals take longer than an order for cards. Deploying the open models on-premise with controlled access, on a Kubernetes-based MLOps platform, is part of our Private AI/ML service, with engineering by our partner Vixen.UNO.

FAQ