Spotted on DEV·TV: FLUX 3 Image Klein just landed, and it renders text like a sign painter

#ai #webdev #showdev #opensource

Tonight a new model popped up on the NEW MODELS channel of DEV·TV, the TV-style browser app I built for developer content. It was FLUX 3 Image Klein from Black Forest Labs, released today (Oct 15, 2026) on OpenRouter.

I went digging. Here's what it is, what it costs, and why it's worth a look.

What is it?

FLUX 3 Image Klein is the small, fast member of Black Forest Labs' new FLUX 3 Image family:

Open weights. You can download it, run it yourself, and fine-tune it.
5.6B parameters. Small enough to fit on a single serious GPU.
Text-to-image and editing. It handles multi-reference edits that keep characters consistent across images, plus style transfer from reference images.
Text rendering. This is the headline feature, and the demo shows why.

The demo that sold me

The OpenRouter playground ships with an example prompt: an editorial photo of a neighborhood storefront at blue hour, with a hand-bent neon sign that reads exactly "OpenRouter".

The result is a 4K, 16:9 image (5456×3072). The cursive neon is spelled correctly, the glow is warm, there are reflections on the wet pavement, and plants fill the windows. It took about 75 seconds and cost $0.036 at that 4K tier.

Not long ago, getting an image model to spell a word right was a coin flip. Now it's the selling point.

What does it cost?

Pricing starts at about $0.008 per image at lower resolutions, and higher resolution tiers cost more. For comparison, the flagship FLUX.3 Image starts at about $0.041 per image, so Klein is the "generate a lot, cheaply" option. It fits thumbnails, product shots, and title cards.

Calling it from code

OpenRouter exposes an Images API. In Python:

from openrouter import OpenRouter
import os

with OpenRouter(api_key=os.environ["OPENROUTER_API_KEY"]) as client:
    res = client.images.generate(
        model="black-forest-labs/flux-3-image-klein",
        prompt='A vintage CRT TV glowing on a desk at night, screen reads "DEV·TV"',
        aspect_ratio="16:9",
        resolution="2K",
        seed=42,
    )

Useful knobs:

aspect_ratio and resolution set the output shape and size.
seed gives reproducible results.
n asks for up to 10 images.
input_references handles image-to-image edits.

Prompting tip: write full descriptive sentences, not keyword soup, and put any exact text in quotes. The model treats quoted text as "render this literally."

How DEV·TV caught it

DEV·TV is a single index.html on GitHub Pages with no backend. The NEW MODELS channel pulls fresh listings straight from OpenRouter, using a 24-hour window so release-day drops like this one show up on screen while they're still news.

Other channels run alongside it:

AGENT SKILLS surfaces trending Claude skill repos.
Spotlight rotates indie dev projects.

Open it in a tab, put it on a second monitor, and new models come to you.

👉 Watch it live: https://shouvik12.github.io/devtv

What would you generate first with a model that can actually spell? Drop your prompt in the comments 👇

DEV Community