How much VRAM does FLUX 3 Image Klein need?

FLUX 3 Image Klein is an open-weight image model, not a chat LLM. The distilled BF16 generator file is 11.1 GB, the FP8 variant 6.3 GB and the NVFP4 variant 5.3 GB; the text encoder (8.9 GB in two shards) and VAE (0.3 GB) load on top. Black Forest Labs says an INT8 build runs on GeForce RTX cards with as little as 8 GB of VRAM, but the file listing we checked on 16 October 2026 does not contain an INT8 filename.

So the 8 GB figure depends on that INT8 build. The files you can download now are bigger: the BF16 generator, text encoder and VAE add up to about 20.3 GB, before any working memory. We haven’t measured a Klein run on any GPU.

Key specs

GB here means 10⁹ bytes and GiB means 2³⁰ bytes. GPU sizes on this site are in GiB; see why file size isn’t the VRAM you need.

Weight files (checked 16 October 2026)

The distilled repo has the generator twice: as a single file and as a two-shard transformer. Those are two copies of the same thing, so don’t add them together. The text encoder is separate. ae.safetensors and vae/diffusion_pytorch_model.safetensors differ by 2,264 bytes, so both are listed.

The base repo uses the same text encoder and VAE sizes. Its generator files:

The distilled single-file generator, both text-encoder shards and the VAE file add up to 20,309,337,412 bytes (20.309 GB, 18.914 GiB). That leaves out ae.safetensors, the transformer shards and the FP8/NVFP4 variants. It’s a sum of files, not a measured load; activations and diffusion working memory come on top.

The blog’s 11.2 GB (BF16) and 5.6 GB (FP8) figures are rounded publisher numbers and don’t match the 11.097 GB and 6.271 GB files exactly. No filename containing int8 appeared in the 42-file distilled listing or the 39-file base listing.

What Black Forest Labs says about hardware

The INT8 build runs on GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. NVIDIA’s timings, as reported in the same post: about 4 seconds per megapixel on an RTX 5090 and under 9 seconds on an RTX 4070 for the distilled INT8 build in ComfyUI, and under 0.6 seconds per megapixel image on GB300 with NVFP4. We haven’t reproduced any of these.

The release notes say both variants run on consumer GPUs down to an RTX 3090. Cards in the classes they name: RTX 4060 8GB (see what fits in 8 GB), RTX 4070 12GB, RTX 3090 24GB and RTX 5090 32GB. None of these pages has a measured Klein run.

The release notes also link a separate ComfyUI weight repo, which we haven’t summed.

API price

Per image, same for text-to-image and editing (release notes, 15 October 2026):

The VRAM calculator and KV cache calculator cover open language models such as Qwen3.8-27B, not diffusion models. Two GPUs don’t pool into one bigger card; see two GPUs and unified memory. File sizes above are from the publisher’s listings, not our measurements. How we calculate.

Sources last checked 2026-10-16.