FLUX 3 Image Klein VRAM calculator

Can FLUX 3 Image Klein run on 8, 12, 16 or 24 GB?

BFL’s release post reports an INT8 setup running through ComfyUI or Diffusers on GeForce RTX GPUs with as little as 8 GB. Its 1024×1024, 16-step examples take about four seconds on an RTX 5090 and under nine seconds on an RTX 4070. These are publisher-reported results for those implementations, not measurements by ModelVRAM.

The model card’s reference-code measurements use a B200 at 1024×1024 with the prompt upsampler disabled: 31 GB in BF16, 26 GB in FP8 and 25 GB in NVFP4. They describe a different runtime. A 24 GB card is below each of those reported peaks; an optimized consumer-GPU workflow can still use less memory. Match the build, encoder placement and input size before comparing either set of numbers with your card.

Transformer files versus the whole pipeline

The table adds the root package’s two BF16 text-encoder shards (8,875,719,344 bytes) and its image autoencoder (336,211,292 bytes) to each transformer. These are exact download sizes from the publisher’s file listing, checked October 15, 2026. The alternative Diffusers transformer and VAE paths are another packaging of the same components; do not add both copies.

GiB = 1,073,741,824 bytes. File totals are storage footprints, not peak VRAM. Quantizing only the transformer leaves the BF16 encoder unchanged. Activation tensors, image decoding, CUDA allocations and any prompt upsampler add memory; CPU offload can keep some weights out of GPU memory.

The ComfyUI INT8 build has its own component files

The official ComfyUI package contains a distilled flux3_image_klein_sd_int8_convrot.safetensors transformer of 5,827,874,779 bytes (5.43 GiB), an INT8 Qwen3-VL-4B encoder of 4,858,559,240 bytes (4.52 GiB), and the 0.31 GiB VAE. Downloaded together they total 10.27 GiB. That exceeds 8 GiB as files, so the low-VRAM result cannot be explained by keeping all those components resident simultaneously at their file sizes.

Use the publisher’s workflow for its component loader and memory handling. More reference images, larger output dimensions and concurrent jobs can increase the working set. A weight-only “fits” label does not establish that a particular edit or generation will finish on the card.

Distilled, Base and prompt upsampling

The distilled model targets generation and editing with 16 steps. Klein Base is the undistilled training and research checkpoint; its card specifies 50 steps and guidance 3.5. The package contains separate encoders and a VAE as well. Neither variant is a chat language model, so this page does not assign it a language-model KV cache.

The reference implementation normally runs a separate prompt upsampler. The model card reports about 28 GB of additional GPU memory for that model on an 80 GB GPU; it is excluded from the 31/26/25 GB figures above. The image pipeline takes a detailed scene caption. Supply that caption from a separately run upsampler when the generation workflow does not load one.

Sources and next steps

The weights use the FLUX Kommunity License; review the linked terms for your use. Download access is gated on Hugging Face. This page uses public metadata and official reports and has not run a GPU benchmark.

For other diffusion pipelines, use the image and video VRAM calculator, which separates transformer, encoder and VAE placement. The Qwen-Image 2.1 calculator includes a separate set of published files and measured working-memory ranges. See how ModelVRAM distinguishes file sizes, estimates and measurements.