Back to blog

FLUX 3, Filed: The Family Map, the Open-Weights Klein, and What to Do This Week

The FLUX 3 release, sorted: one multimodal family, Klein as the open-weights 5.6B image model (repo timeline Oct 13–15, 2026), Max as a partner-platform tier — plus who should act now and who should wait, with sources.

Oct 22, 2026Iris Novak avatarIris Novak

Virtual editorial column · Model desk: releases, specs, and costs

MediaMuse editorial content; the profile image is AI-generated and does not represent a real customer, independent reviewer, or named human expert.

AI-assisted filing; specs below are taken from official Black Forest Labs pages and repositories, checked on the day of publication. Anything not tested in-house is labeled as such.

A landscape-format generated image used for layout planning

Black Forest Labs shipped FLUX 3 as one family, and the internet immediately flattened it into a product name you either have or don't. That flattening is where the confusion lives, so let me file the actual structure first, then get to the part that matters: what an image workflow should do about it this week. Short answer at the bottom, as always, and it is checkable.

The family, drawn properly

FLUX 3 is described by BFL as one multimodal model covering image, video, audio, and action prediction — not four separate products. Within it, the two names everyone keeps colliding:

  • FLUX 3 Image Klein — the open-weights image release: a 5.6-billion-parameter image transformer that does text-to-image, single-reference editing, and multi-reference composition (up to ten reference images in one prompt) in one model. The text encoder is counted separately: Qwen3-VL-4B-Instruct. Gated download, Hugging Face and NVIDIA Build, and — the fine print that decides most "can I use this commercially" questions — free commercial use limited to smaller businesses, per the licensing terms as reported.
  • FLUX 3 Max — a tier, not a model you download: BFL's top performance level, currently served through partner platforms (fal, Krea). The headline specs from the official pages: a 10-second video generated in roughly 12 seconds, text or image to start, up to 20 seconds at 720p/1080p with synchronized audio.

Calling both "FLUX 3" is technically true and practically useless. Klein is a self-host story. Max is a hosted-video story. Almost nobody's decision overlaps.

The repo timeline, because "how new" is a real question

For anyone gauging freshness — this matters for whether tooling has settled:

  • Oct 13, 2026 — the official BFL repository for klein appears.
  • Oct 14 — the official ComfyUI support repository follows.
  • Oct 15 onward — community GGUF and quantization repositories appear densely.

We track this timeline on our full Klein page — specs, both checkpoints, the gated download, the ComfyUI route, and the FLUX.2 comparison. That page is unaffiliated with BFL and links only official sources; this column adds the opinion layer on top.

The two hidden costs in the open-weights story

I run a model desk, so I read release notes the way other people read ingredient lists. Two things about klein that the announcement coverage mostly skipped:

It needs a prompt upsampler. Short prompts used directly give generic output — the official route is flux-3-image-klein-pup, or any LLM that expands your prompt into a full-scene caption. That is an extra step, and if you are self-hosting, an extra component. Not tested in-house; taken from the official repository's own guidance.

The interesting capability is positional, not stylistic. Klein's highlighted feature is bbox prompting — placing elements with boxes on a 0–1000 grid, then recoloring, moving, replacing, or removing them by changing the prompt. That is layout control, the thing stock prompt-and-pray generators don't give you. For product compositions and reference-driven edits, that is the spec line worth remembering.

What to actually do this week

Split by who you are, no hedging:

  • Self-hosting, comfortable with ComfyUI — the official support repo landed Oct 14 and quantization traffic started Oct 15. You are the target user. Budget for the upsampler step.
  • Using hosted generators (like this site's) — klein changes nothing for you yet. Hosted image workflows are still served better by the models already on the menu — for example, Ideogram V3's ratio and edit modes at 4–10 credits an image, or the browser generator itself. Watch this space rather than switching.
  • Video-curious — Max is a fal/Krea tier with synchronized audio. If you were waiting for BFL video, the partner pages are the price list; we do not serve it and have no plans to claim otherwise.

And the checkable call: if BFL or a partner publishes hosted klein API pricing within 30 days, hosted generators will add it within the week — that is how the Ideogram and Nano Banana cycles went. If nothing appears by late November, klein stays a self-host story into 2027, and the naming confusion becomes the only thing most people remember about the release. Come back and check.

Sources: Black Forest Labs — FLUX 3 model pages (multimodal family; Max tier via fal/Krea; 10s-in-~12s spec) · FLUX 3 Image Klein on Hugging Face (5.6B transformer; Qwen3-VL-4B-Instruct encoder; gated download) · VentureBeat on the Klein open-weights release (licensing: free commercial use limited to smaller businesses) · repo timeline Oct 13–15, 2026, via official repositories.

Related guides