FLUX.2 Klein 4B and 9B: Verified Model and VRAM Facts

By Dominic Marrocco, 42 UK Research. Technical claims below are bounded to the linked primary sources.

Evidence review: 5 August 2026

What the official sources support

This page replaces an earlier, unsupported two-gigabyte hardware claim. The legacy URL is retained so existing links still work, but neither official Black Forest Labs model card cited here supports a 2 GB requirement. The 4B base card says the model fits in about 13 GB of VRAM; the 9B card says its model fits in about 29 GB of VRAM. Both are publisher specifications for the configurations they describe, not independent 42 UK measurements.

FLUX.2 klein base 4B

Black Forest Labs identifies FLUX.2 klein base 4B as a four-billion-parameter base model for image generation and editing. Its official card publishes the weights under the Apache 2.0 licence and states that the model can run with as little as 13 GB of VRAM. “Base” matters: this is the undistilled variant described by that card, so settings from a distilled release should not be copied across without checking the model-specific instructions.

FLUX.2 klein 9B

The official 9B model card describes a nine-billion-parameter flow model paired with a Qwen3 eight-billion-parameter text embedder. It describes the release as distilled for four-step generation, says it fits in about 29 GB of VRAM, and places it under the FLUX non-commercial licence. That licence is materially different from Apache 2.0; anyone planning commercial use should read the current licence terms at the official source rather than infer permission from the shared FLUX.2 klein name.

Verified comparison

• Published variant: 4B is a base model; the cited 9B release is distilled.
• Published licence: the cited 4B base weights use Apache 2.0; the cited 9B weights use a non-commercial licence.
• Published 4B memory statement: the 4B card says “as little as 13 GB of VRAM” for its described setup. This is a publisher statement, not an independently reproduced 42 UK result.
• Published 9B memory statement: the 9B card says the model fits in about 29 GB of VRAM. This is likewise attributed publisher evidence.
• Not established here: which variant is faster, which produces better images, or the minimum memory for every precision, resolution, operating system and offload configuration.

Using either model with ComfyUI

ComfyUI’s official repository documents supported installation paths for Windows, Linux and macOS. Check the current system requirements before choosing a Python, PyTorch or hardware path; those requirements can change independently of this article. Obtain weights and workflow instructions from the relevant Black Forest Labs model card, not from an unverified mirror.

If memory allocation fails, work through the official troubleshooting guide. It lists mitigations including --lowvram, reducing resolution or batch size, disabling previews with --preview-method none, and adjusting reserved VRAM. These controls can reduce memory pressure, but they do not prove that a specific model works in 2 GB.

A decision process that does not invent results

1. Choose the licence that fits the intended use, using the current official model card.
2. Compare the publisher’s hardware statement with the actual available system RAM and VRAM.
3. Follow the current ComfyUI installation and model-specific workflow instructions.
4. Record model revision, precision, workflow, resolution, batch size, command-line flags and peak memory before publishing any benchmark claim.

For reproducible site evidence, continue to the low-VRAM benchmark methodology, the installation matrix, and the validated workflow library.