FLUX.1 [dev] VRAM: how big are the model files?
Short answer: the files Diffusers loads for FLUX.1 [dev] add up to 33.7 GB on Hugging Face.[1] That is weights only: it does not include the memory used while an image is generated (activations), which depends on resolution, batch size and settings. The Hugging Face Diffusers docs say "Flux is a very large model and requires ~50GB of RAM/VRAM to load all the modeling components."[2] We don't publish our own activation estimate, because we couldn't find a sourced figure for it.
FLUX.1 [dev] is Black Forest Labs' 12B-parameter text-to-image model. Hugging Face lists 11,901,408,320 parameters for the transformer, the repo was created on 2024-07-31, and it is released under the FLUX.1 [dev] Non-Commercial License (the repo is gated: you accept the license before downloading).[1] Hugging Face counts about 542,979 downloads in the last 30 days (checked November 3, 2026).[1]
Weights only: file sizes on Hugging Face
Exact sizes from the black-forest-labs/FLUX.1-dev file listing, in Diffusers layout (1 GB = 1,000,000,000 bytes; split files added together; checked November 3, 2026).[1]
| Component | Files | Size |
|---|---|---|
| Transformer (12B, BF16) | transformer/ (3 files) |
23.80 GB |
| T5-XXL text encoder | text_encoder_2/ (2 files) |
9.52 GB |
| CLIP text encoder | text_encoder/model.safetensors |
0.25 GB |
| VAE | vae/diffusion_pytorch_model.safetensors |
0.17 GB |
| Total (weights only) | 33.74 GB |
The repo also has a single-file flux1-dev.safetensors (23.80 GB, the transformer alone, used by ComfyUI and similar tools) and ae.safetensors (0.34 GB, the VAE).[1] You need one copy of each, not both layouts.
Weights only vs. what you need to run it
| Figure | Size | What it covers | Source |
|---|---|---|---|
| Model files (this page) | 33.7 GB | All four components' weights, no activations | Hugging Face file listing[1] |
| Diffusers docs | ~50 GB | "RAM/VRAM to load all the modeling components" | Diffusers Flux docs[2] |
So a 24 GB card can't hold every component at once at full precision. The documented ways around that, such as CPU offload, FP8 and 8-bit loading, are covered in our FLUX.1 VRAM guide.
Sources
All sources checked on .
- Hugging Face, black-forest-labs/FLUX.1-dev (file listing with exact sizes, parameter count, license, repo date, 30-day downloads; checked November 3, 2026). https://huggingface.co/black-forest-labs/FLUX.1-dev/tree/main
- Hugging Face Diffusers, Flux pipeline docs, "Optimize" (checked November 3, 2026). https://huggingface.co/docs/diffusers/api/pipelines/flux#optimize