MediaMuse AI
Credits—

FLUX 3 Klein: What It Is, Specs, and How to Run It | MediaMuse AI

FLUX 3 Klein is the new open-weights 5.6B image model from Black Forest Labs: key specs, both checkpoints, the gated weight download, the ComfyUI route, and the FLUX.2 klein comparison.

Last updated: 2026-10-15

FLUX 3 Image Klein is the open-weights image model Black Forest Labs released in October 2026: a 5.6-billion-parameter image transformer that generates and edits images in one model. MediaMuse is not affiliated with Black Forest Labs, does not host the model, and this page is not a review — it collects what the official release states, points only at official sources, and describes the browser-based image generator we do publish.

What is FLUX 3 Image Klein?

It is the open klein release of the FLUX 3 image generation line. The 5.6B parameter count is the image transformer alone — the text encoder is counted separately and is Qwen3-VL-4B-Instruct. One model covers three jobs: text-to-image, editing from a single reference image, and composition from multiple references, with up to ten reference images carried in a single prompt.

The capability the release highlights as its own is bbox prompting: you place elements with boxes on a 0–1000 grid inside the prompt, then recolor, move, replace, or remove those elements by changing the prompt. Combined with the reference-image limit, that is what makes the editing story more than a caption-and-hope workflow.

The release timeline, for anyone gauging how new this is: the official BFL repository was created 2026-10-13; flux-3-image-klein-comfy, the official ComfyUI support repository, followed on 10-14; and community GGUF and quantization repositories started appearing densely from 10-15.

Key specs

  • Model size — 5.6B parameters (image transformer; the text encoder is counted separately)
  • Text encoder — Qwen3-VL-4B-Instruct
  • Capabilities — text-to-image, single-reference editing, and multi-reference composition in one model
  • Reference images — up to 10 per prompt
  • Klein checkpoint — distilled: 16 steps, no guidance; built for self-hosted generation and editing
  • Klein Base checkpoint — undistilled bf16: 50 steps, guidance 3.5; built for fine-tuning, LoRA, and research
  • Prompting — a prompt upsampler is required: the official flux-3-image-klein-pup, or any LLM that expands the prompt into a full-scene caption. Short prompts used directly give generic output
  • Diffusers — cannot run the model standalone
  • Layout control — bbox prompts on a 0–1000 grid, then recolor, move, replace, or remove via the prompt
  • License — flux-kommunity-license; the Hugging Face repository is gated, so the conditions must be accepted before the weights can be downloaded

How to run FLUX 3 Klein locally

The honest version, in the order the official pieces exist. Nothing here is hosted or verified by us — each step names its official source.

  1. Get the weights. The Hugging Face repository is gated under the flux-kommunity-license: accept the conditions on the Black Forest Labs organization page before anything can be downloaded.
  2. Use ComfyUI. flux-3-image-klein-comfy is the official ComfyUI support repository (created 2026-10-14, listed from the Black Forest Labs GitHub organization) — the supported workflow route, rather than a hand-built pipeline.
  3. Upsample every prompt. The model is built around full-scene captions: run prompts through the official flux-3-image-klein-pup upsampler, or through any LLM that expands them. Short prompts used directly give generic output, and Diffusers cannot run the model standalone — there is no shortcut that skips this step.
  4. Pick the checkpoint that matches the job. Klein, the distilled one, runs 16 steps with no guidance and is the checkpoint for generation and editing. Klein Base, the undistilled bf16 one, runs 50 steps at guidance 3.5 and is the one for fine-tuning, LoRA, and research.

Community GGUF and quantized builds have been appearing since 2026-10-15, which is usually how open models get fitted onto smaller machines — read each repository's own page before trusting one. We link only the official sources above.

Generate online instead

If the local route is more setup than the image is worth and you just need a picture today: MediaMuse does not host FLUX 3 Klein. Our AI image generator runs in the browser on other models — and if your actual job is editing, image-to-image takes up to ten reference images on Nano Banana 2.1, with output up to 4K. Terms stated as plainly as the local setup: every run bills per image, the composer shows the exact credit estimate before the run starts, and there is no free tier. Full rates on the Nano Banana 2.1 model page and pricing.

Klein vs FLUX.2 Klein

The two generations share the klein name, and the older one owns most of the existing search traffic — so here is the split, stated only as far as the official pages go.

  • FLUX.2-klein — the January 2026 generation, released as a 4B/9B series. FLUX.2-klein-4B counts about 420,000 cumulative downloads, which is the size of the audience that already arrives knowing the klein name.
  • FLUX 3 Image Klein — the October 2026 release: a 5.6B image transformer with Qwen3-VL-4B-Instruct as the text encoder, up to ten reference images per prompt, text-to-image and editing in one model, and bbox prompting on a 0–1000 grid.

What that means practically: the checkpoints, the step counts, and the prompting rules are all different between the generations, so any tutorial you follow should name the generation it was written for. A page written for FLUX.2-klein in January is describing a different setup than the one the specs above define.

API access

Black Forest Labs serves the distilled Klein checkpoint — the same 16-step, no-guidance checkpoint described above — through its own API, under its own terms and prices on its official site. That is the official route if you want the model without maintaining local weights.

What we publish is different, and we keep the two claims separate: MediaMuse does not integrate the BFL API and does not host or proxy Klein in any form. Our generator runs other models in the browser, billed per image, with the cost shown before every run — try the AI image generator if that is the shape of workflow you came here for.

FAQ

Can I try FLUX 3 Klein online?

Not on this site, and we will not write copy that implies otherwise: MediaMuse does not host the model. The official online route is the BFL API, which serves the distilled checkpoint. If what you actually want is to generate an image online right now, our AI image generator runs in the browser on other models, billed per image with the estimate shown before the run.

Where do I download the weights?

On Hugging Face, in the gated Black Forest Labs repository, after accepting the flux-kommunity-license conditions. Nothing on this site hosts, mirrors, or resells them.

Is there an API for FLUX 3 Klein?

Yes — the official BFL API runs the distilled Klein checkpoint (16 steps, no guidance; the same checkpoint sold for self-hosting). Terms and pricing live on the official site, and we neither resell nor proxy it.

How is FLUX 3 Klein different from FLUX.2 klein?

Generation and size. FLUX.2-klein shipped in January 2026 as a 4B/9B series — FLUX.2-klein-4B sits at about 420,000 cumulative downloads. FLUX 3 Image Klein is the October 2026 release: a 5.6B image transformer, Qwen3-VL-4B-Instruct text encoder, up to ten reference images per prompt, and text-to-image plus editing in one model. The comparison section above has the full split.

How do I deploy it — does Diffusers work?

Diffusers cannot run the model standalone. The supported local route is ComfyUI with the official flux-3-image-klein-comfy repository, the gated weights, and a prompt upsampler in front of every run; community quantized builds exist for smaller setups.

What hardware does it need?

Only the official model card can answer that with authority, and we do not invent VRAM numbers. What the release itself establishes: the model is a 5.6B image transformer plus a 4B text encoder, in two checkpoints, and community GGUF and quantized builds have existed since 2026-10-15 — quantization being the usual route for fitting open models onto smaller cards. Read the requirements on the gated model card before buying anything.

Why do my results look generic?

Because the model is built for full-scene captions, and a short prompt used directly gives generic output — the release states this plainly. Run your prompt through the official flux-3-image-klein-pup upsampler, or through any LLM that expands it into a full-scene caption, before it reaches the model.

Keep reading