FLUX.2 [dev] VRAM Requirements
FLUX.2 [dev] needs about 31.0 GB of VRAM with FP8 weights at 1024×1024, loading one model at a time. In FP16 it needs 60.8 GB (108 GB with everything loaded), and GGUF Q4 brings it to 23.6 GB. The smallest consumer GPU that runs it is the RTX 3090 24GB with Q4 weights, or the RTX 5090 32GB in FP8.
31.0 GB peak VRAM
FLUX.2 [dev] at 1024×1024 with FP8 weights, loading one model at a time (ComfyUI default on smaller cards). Keeping everything loaded needs 56.3 GB.
- Diffusion model 29.8 GB
- Text encoders 22.4 GB
- VAE 0.2 GB
- Peak activations 3.1 GB
- Overhead 0.8 GB
GPU fit
- RTX 4060 8GBToo small
- RTX 3060 12GB36% on GPU
- RTX 4060 Ti 16GB49% on GPU
- RTX 4090 24GB73% on GPU
- RTX 5090 32GBTight fitwith swapping
- RTX 6000 Ada 48GBFitswith swapping
- RTX PRO 6000 Blackwell 96GBFitsall loaded
- Apple M4 Max 128GBFitsall loaded
The VAE decode is the peak here. Turn on tiled VAE decoding to cut it to about 0.8 GB.
FLUX.2 [dev] VRAM by precision
At 1024×1024. Text encoders in FP8 except for the FP16 row.
| Weights | Diffusion model | Peak (swapping) | All loaded | Smallest tier |
|---|---|---|---|---|
| FP16 / BF16 | 59.6 GB | 60.8 GB | 108 GB | 80 GB |
| FP8 | 29.8 GB | 31.0 GB | 56.3 GB | 48 GB |
| GGUF Q8_0 | 31.7 GB | 32.8 GB | 58.1 GB | 48 GB |
| GGUF Q5_K_S | 20.6 GB | 23.6 GB | 47.0 GB | 32 GB |
| GGUF Q4_K_S / NF4 | 17.0 GB | 23.6 GB | 43.4 GB | 32 GB |
VRAM by resolution
FP8 weights, with model swapping.
| Resolution | Latent tokens | Activations | Peak VRAM |
|---|---|---|---|
| 1024×1024 | 4,096 | 0.4 GB | 31.0 GB |
| 1536×1536 | 9,216 | 0.8 GB | 31.4 GB |
About FLUX.2 [dev]
A 32B image model with built-in editing and multi-reference support. Its text encoder is a full 24B language model (Mistral Small 3.1), so the encoder alone is about 45 GB in BF16. On consumer cards it runs with quantized weights (FP8, GGUF or 4-bit) and model swapping, and many setups run the text encoder in system RAM or remotely.
| Component | Size |
|---|---|
| Diffusion model | 32B parameters (Rectified-flow transformer) |
| Text encoders | Mistral Small 3.1 24B (24B) |
| VAE | 0.084B parameters |
| Classifier-free guidance | Distilled (1 pass per step) |
| Released | 2025-11 |
| License | FLUX Non-Commercial License |
FLUX.2 [dev] VRAM: common questions
Can FLUX.2 dev run on an RTX 4090 or 5090?
On a 32 GB RTX 5090, yes with FP8 or GGUF weights and model swapping. On a 24 GB RTX 4090 it works with a GGUF Q4 transformer (about 17 GB) plus a GGUF text encoder, or with the text encoder offloaded to system RAM; the FP8 text encoder alone is about 22 GB.
Why is FLUX.2 so much bigger than FLUX.1?
The transformer grew from 12B to 32B, and the T5 text encoder was replaced with a 24B Mistral language model. Together they need over 100 GB in BF16 if loaded at once, which is why quantization and model swapping are essential.
How much VRAM does FLUX.2 [dev] need?
At its default 1024×1024: about 60.8 GB in FP16 with model swapping (108 GB with everything loaded), 31.0 GB in FP8 and 23.6 GB with GGUF Q4 weights.