Datasheet No. 01 · Local AI on your own GPUCalculated from 839 real model files

Will it run
on your GPU?

FLUX, Qwen-Image, Krea 2, Wan, LTX, MiniMax H3 and 43 more. Which file to download, how much VRAM it really needs, and what your card can handle. Every number says where it came from.

I have Your GPU and I want to run AI model.

…

Loading the answer.

Fig. 1 — LiveStep 01/20σ 1.000Seed —

Memory map · RTX 5060 Ti 16 GB · FLUX.1 dev FP8 · Sampling14.2 GB / 16 GB

08 GB16 GB

Weights FP8 · 11.9 GBWorking memory · 1.5 GBReserve · 0.8 GBFree · 1.8 GB

49image & video models

98desktop & laptop GPUs

839model files, sized to the byte

2453answer pages

01The memory ruler

Every model on one scale. Drag the red line to your VRAM.

Each dot is one file of a model, placed at the VRAM it needs. Everything left of the line fits.

Your VRAM

16GB

0 of 49 models fit entirely.

VRAM in GB

Model

081624324048

Verdict · file

Image models

Video models

A file of this model (VRAM it needs)Fits in your VRAMBest file that fitsYour VRAM

02Asked most

Browse by VRAM

The shaded part of each column is how many models fit at all.

6GB7 run well
12 fit at all 8GB8 run well
22 fit at all 10GB12 run well
29 fit at all 12GB17 run well
37 fit at all 16GB25 run well
47 fit at all 20GB32 run well
48 fit at all 24GB43 run well
49 fit at all 32GB47 run well
49 fit at all

04Tools and special cases

Plan the setting, the upgrade, the laptop

CalculatorBigger images, longer clips, LoRAs and ControlNetsSet resolution, frames and batch; see which file still fits → UpgradeWhat would a new GPU change?Your card vs the one you want: what newly runs well → TrainingLoRA training: how much VRAM49 figures from the trainers' own docs, with settings → LaptopsLaptop GPUs have less VRAM than their names suggestThe RTX 4090 Laptop has 16 GB, not 24. 32 laptop GPUs → MacsApple Silicon: what fits, what works, how slow16 to 512 GB, M1 to M5. FP8 files do not help on a Mac → Compare44 GPU matchups, model by modelRTX 3090 vs 5060 Ti, 7900 XTX vs 4090… →

05No guesses without a label

Where every number comes from

How the numbers work →

calculated

From real file sizes

Every model file is sized to the byte from Hugging Face. Add the model's working memory and a small system reserve, and you get the verdict of what fits entirely.

measured

On my own GPU

I make AI images with FLUX in ComfyUI on an RTX 5060 Ti 16 GB. A benchmark script times each model and logs peak VRAM; those results are being added.

reported

From other people

69 real-world results from GitHub, Hugging Face and first-hand blogs, copied as published with a link to each. See them all or send yours.

06Newest first

Recent models

Month by month, with email alerts →

2026-09Qwen-Image 2.17.1B · image · fits from 6 GB 2026-07Mage-Flow (Microsoft)4.12B · image · fits from 6 GB 2026-07Wan Animate 2 (14B)14B · video · fits from 12 GB 2026-07LTX-2.5 (22B)22B · video · fits from 14 GB 2026-07MiniMax H3 (33B)33.1B · video · fits from 22 GB 2026-07MiniMax H3 Prunedpruned 33B · video · fits from 15 GB 2026-06Krea 2 (Turbo)12.8B · image · fits from 8 GB 2026-06Ideogram 49.3B · image · fits from 15 GB

07Short and practical

Guides

G.01GGUF Q8, Q6, Q5, Q4: which one should you download?Files G.02FP8, GGUF or 16-bit in ComfyUI: what is the difference?Files G.03How much VRAM do you need for AI images and video in 2026?Buying G.04Out of memory in ComfyUI: fixes in the order to try themFixes G.05Text encoders: why T5, UMT5 and Qwen-VL eat your memoryExplainer G.06How much system RAM do you need for ComfyUI?Buying G.07AMD and Intel GPUs for ComfyUI in 2026GPUs G.08Local AI on a laptop: what the GPU name does not tell youGPUs G.0996 GB of "VRAM" on Strix Halo: what fits, and what it costs youGPUs G.10Why video models need so much more memory than image modelsExplainer G.11INT8, NVFP4, MXFP8: the new model file formats of 2026Files

Will it runon your GPU?