Artificial intelligence news.
Straight from the original announcements.

Open video models range from 14GB to 80GB of VRAM

November 16, 2026By the AI News Bites Editorial TeamSource: LTX (Lightricks)
Open video models range from 14GB to 80GB of VRAM
Image: LTX (Lightricks)

Published VRAM requirements for open video models range from 14GB to 80GB, according to an LTX comparison. The figures are recipe-specific, because offloading, data types, resolution and frame count all affect how much GPU memory generation requires.

Key points
  • HunyuanVideo-1.5 lists a 14GB floor with model offloading.
  • Wan2.2 TI2V-5B lists 24GB with offload flags.
  • LTX-2.5 lists 32GB minimum and recommends 80GB GPUs.
  • Wan2.2 T2V-A14B lists 80GB for its single-GPU command.
  • FP8, offloading and the conv VAE can reduce LTX-2.5 memory use.

Published VRAM floors are not directly comparable

LTX lists four published single-GPU requirements. HunyuanVideo-1.5, an 8.3B model, has a 14GB minimum when model offloading is enabled. Wan2.2 TI2V-5B lists 24GB, with the requirement measured using model offloading, parameter dtype conversion and a T5 text encoder running on the CPU.

LTX-2.5, a 22B model, lists 32GB as its minimum, while the company recommends an A100 with 80GB or an H100. Wan2.2 T2V-A14B, with 27B total parameters and 14B active, lists 80GB for its single-GPU command. These figures support 480p, 720p or higher outputs depending on the model and configuration, but they are not benchmark results conducted under identical conditions.

LTX says the lower floors generally make a model possible to run, rather than fast to run. Offloading moves weights between system RAM and the GPU, which can reduce GPU memory use while slowing generation. LTX-2.5 lists 32GB of system memory as a minimum and 64GB or more as recommended.

Resolution and frame count drive memory use

For a video diffusion transformer, sequence length is determined by resolution and frame count. The LTX trainer gives the default LTX VAE formula as sequence_length = (H / 32) * (W / 32) * ((F - 1) / 8 + 1). A 768×448×89 clip produces 24 × 14 × 12, or 4,032 tokens.

Increasing both width and height multiplies the spatial component by four, while frame count grows the sequence roughly linearly. The guide says every eight additional frames adds one latent frame. LTX-2.5 also requires frame counts that satisfy num_frames % 8 == 1, with valid examples including 97 and 121, and dimensions divisible by 32.

Token counts cannot be compared directly between models. Wan2.2 TI2V-5B uses a different VAE compression scheme, with a total ratio of 4×32×32 after patchification. That difference is another reason the models' headline VRAM figures do not line up.

LTX-2.5 offers several ways to reduce GPU memory

LTX's documentation recommends running the transformer in FP8 and enabling CPU offload. The fp8-cast option downcasts a bf16 transformer on the fly, while fp8-scaled-mm is intended for a native FP8 checkpoint on Hopper and newer GPUs.

CPU offload streams weights from system RAM per layer. Disk offload reduces memory further, but reads weights from storage on demand and is slower. LTX also describes a conv VAE as lighter and faster than the default diffusion decoder, while the distilled pipeline uses eight predefined sigmas as a recommended fast path. The company recommends starting with FP8, adding CPU offload, switching to the conv VAE and using disk offload only when necessary.

The LTX-2.5 quick-start download for the distilled split pack is roughly 66 GiB. The model also includes a Gemma 4 12B text encoder, and LTX says loading that encoder in 8-bit saves about 8 GiB of VRAM while caption embeddings are computed.

What the figures mean for GPU owners

LTX's tier map places HunyuanVideo-1.5 within reach of a 16GB GPU when offloading is used. A 24GB card, such as an RTX 4090 class card, can run HunyuanVideo-1.5 and Wan2.2 TI2V-5B at 720p with the listed offload flags. A 32GB card adds LTX-2.5 at its documented minimum, while an 80GB A100 or H100 can run all four listed recipes, including the Wan2.2 A14B single-GPU command.

Users whose hardware falls below a model's floor, or who find offloading too slow, can use the hosted LTX API or rent cloud GPUs, according to LTX. The practical choice therefore depends on both available VRAM and the speed required for iteration. For local setups, the source recommends starting with the vendor's own recipe and then applying memory-saving options one at a time.

Related news

LTX compares video API prices, billing rules and self-hosting costs
BusinessNovember 16, 2026

LTX compares video API prices, billing rules and self-hosting costs

LTX has compared per-second pricing for LTX-2.5, Veo 3.1, Gen-4.5, Kling 3.0 and Seedance 2.0, finding that billing rules can matter as much as headline rates. The…

LTX Releases VFX Tools for Mattes, Archival Restoration and HDR
Products and toolsOctober 26, 2026

LTX Releases VFX Tools for Mattes, Archival Restoration and HDR

LTX, the Lightricks platform, has presented a set of VFX tools aimed at the production pipeline, including Alpha Gen for soft mattes, Refine and Restore for archival…

LTX-2.5 Offers Raw Open-Weights Checkpoint for Robotics and Physical AI
ModelsOctober 19, 2026

LTX-2.5 Offers Raw Open-Weights Checkpoint for Robotics and Physical AI

LTX (Lightricks) says its 22-billion-parameter LTX-2.5 model includes a raw pretrained checkpoint meant as a starting point for robotics, autonomous vehicle and other…