The honest answer to "what does LTX 2.5 need to run" is the same shape as it was for 2.3: the model is datacenter-shaped with consumer escape hatches, and Lightricks' official positioning is a 16 GB minimum with the low-memory paths being int8 builds and offloading. Within hours of release, community reports of the int8 path running on 8 GB-class GPUs appeared — preliminary, third-party, and consistent with the files Lightricks shipped.
This guide separates what is documented from what is reported, and maps the official files to the hardware that can actually run them. Sourced from the LTX-2.5 model card, the LTX-2 GitHub README, the ltx.io LTX-2.5 page, and the ComfyUI LTX-2.5 tutorial. Checked August 13, 2026.
The documented baseline
The official ltx.io page lists Min VRAM: 16GB for LTX-2.5. That is the vendor's line for the general case — which means the bf16 distilled transformer (~44 GB on disk, a 22B model) plus the diffusion VAE and encoder, with some form of memory management.
The GitHub README documents the software baseline for the Python path:
- Python >= 3.12, CUDA >= 12.7, PyTorch ~= 2.7 recommended
- The
nattenextra is the fastest backend for the diffusion video decoder — Linux + CUDA only; Windows/macOS fall back to Triton or eager implementations (slower, same install command)
The three official low-memory paths
LTX-2.5 ships three documented ways to cut memory, and they are not interchangeable:
1. ComfyUI int8 builds — the consumer path.
The comfy-int8-convrot variants of both the distilled and dev transformers are the official ComfyUI builds, explicitly labeled "ComfyUI only — not for ltx-pipelines / PyTorch." The official workflows default to the int8 distilled build, which is the configuration behind the day-one 8 GB-class reports. This is the most accessible path for consumer hardware — and the model card is clear that these files are not loadable by the Python path.
2. NVFP4 — the Blackwell path.
ltx-2.5-22b-distilled-transformer-nvfp4.safetensors targets Blackwell GPUs (RTX 50-series and datacenter). It works in ComfyUI, or in ltx-pipelines with --quantization nvfp4-prequant and ltx-kernels. If you have a Blackwell card, this is the documented quality-preserving quantization for it.
3. Runtime FP8 + offload — the ltx-pipelines path.
For the Python stack, the README's memory lever is the classic pair: --quantization fp8-cast (downcasts the bf16 transformer on the fly; works on any FP8-capable GPU) or --quantization fp8-scaled-mm on Hopper+, plus --offload {cpu, disk}. The model card repeats the pattern: "Downcast bf16 transformer on the fly + CPU offload."
The pattern across all three: nobody promises you the bf16 files on 12 GB. The documented consumer route is the int8 ComfyUI path; the Python route is fp8-cast plus offload, which trades speed for fit.
What actually runs, by hardware class
| Hardware | Documented path | Notes |
|---|---|---|
| 8 GB-class (RTX 4060/3060 laptop-class) | ComfyUI int8 distilled (community-reported, preliminary) | Third-party day-one reports; verify on your card |
| 12–16 GB (RTX 4070/4080 class) | ComfyUI int8 distilled; ltx-pipelines fp8-cast + CPU offload | The 16 GB figure is the vendor's stated minimum |
| 16–24 GB (RTX 4090, 3090) | Comfortable int8; bf16 with offload | Diffusion VAE decode is the memory-heavy stage |
| Blackwell 50-series | NVFP4 distilled (official) | The documented quantization path for Blackwell |
| Hopper+ datacenter | fp8-scaled-mm; full bf16 | Where the vendor's 6.8 s / 10 s clip claim was measured (2× GB200) |
The memory profile is not uniform across a generation: the diffusion video decoder is explicitly documented as heavier (more VRAM, longer decode) than the conv VAE — so on tight VRAM, swapping to the conv VAE is a documented lever too (see our VAE guide).
Speed expectations by class
The vendor's speed figures are datacenter numbers: 6.8 s for a 10-second 720p clip on 2× GB200 at steady state, and 23.7 s for their API at 1080p (measured end-to-end on fal.run including queue time, per the official page). None of that transfers directly to consumer cards.
What transfers: the 8-step distilled schedule (8 steps stage 1, 4 stage 2, CFG=1) means wall-clock time scales with your card's throughput on 8 denoising steps — the structure is cheap regardless of hardware, which is why the int8 + distilled combination is the one people report usable times on. The vendor's own framing ("better results, less compute") is built on that combination.
Don't guess whether your GPU is enough — test it. Generate with LTX 2.5 in your browser at ltx23.app — zero downloads, zero VRAM — and evaluate quality and prompt behavior before you download 66 GB to test locally.
FAQ
What is the minimum VRAM for LTX 2.5? The official minimum is 16 GB for the general case. The ComfyUI int8 builds are the low-memory path (community reports of 8 GB-class cards exist as of day one, preliminary); ltx-pipelines offers fp8-cast + CPU/disk offload.
Can I run LTX 2.5 on 8 GB of VRAM? Preliminary community reports say yes via the official ComfyUI int8 distilled build. Not vendor-documented yet — verify on your card.
Is LTX 2.5 ComfyUI-only for low VRAM? The int8-convrot builds are ComfyUI-only by design. The Python path's low-memory route is fp8-cast + offload (or NVFP4 on Blackwell).
What are the software requirements? Python >= 3.12, CUDA >= 12.7, PyTorch ~= 2.7 recommended for ltx-pipelines. ComfyUI needs the latest version for the LTX-2.5 templates.
Does LTX 2.5 run on Mac? The official docs don't claim macOS support; the natten backend (fast diffusion-VAE decode) is Linux + CUDA only, with macOS falling back to slower implementations. No official Apple Silicon path is documented at release.
Bottom line
LTX 2.5's requirement picture: 16 GB official minimum, three documented low-memory paths (ComfyUI int8, NVFP4 for Blackwell, fp8-cast + offload for Python), and a day-one community signal that the int8 ComfyUI path runs on 8 GB-class cards. The speed headlines are datacenter numbers; your real lever is the 8-step distilled schedule. Test the model's quality at ltx23.app before committing to the download, and only then match the file set to your hardware class.
Sources
Verified August 13, 2026 against primary documentation:
- LTX-2.5 official model page (ltx.io) — Min VRAM: 16 GB, speed measurement (6.8 s on 2× GB200), API speed (23.7 s via fal.run)
- Lightricks/LTX-2.5 — Hugging Face model card — int8-convrot (ComfyUI only) and NVFP4 variants, low-VRAM tips (fp8-cast, offload)
- Lightricks/LTX-2 — GitHub README — software baseline (Python 3.12+, CUDA 12.7+, PyTorch 2.7), natten backend, fp8 policies, distillation step counts
- ComfyUI official LTX-2.5 tutorial — int8-convrot builds in official workflows, model storage layout
- LTX-2 optimization documentation (GitHub) — FP8 policies, diffusion VAE decoder backends, gradient estimation
8 GB VRAM reports are preliminary third-party claims as of one day after release, not vendor documentation. Speed figures are vendor-published and measured on datacenter GPUs.
