The FLUX.2 family

A licence that changes with model size

You can see this in Black Forest Labs own table and it gets lost almost everywhere else: klein comes in two sizes under two licences. klein 4B ships under Apache 2.0, so commercial use is free. klein 9B ships under the FLUX Non-Commercial License, which means you may download the same weights but you may not use them for client work. The dev model also has weights, and it is non-commercial too.

Now put the scores next to the licences. On the Arena text-to-image board klein 9B scores 1069.6 and klein 4B scores 1030.1. The version with the freer licence is the weaker model. If your work is commercial the choice was made for you, and you have to live with the 1030. This is the only place in this reference where a legal document reaches directly into your output quality.

The price here is per megapixel, not per image

Google and OpenAI charge one fixed number per size tier. Black Forest Labs does not: the cost climbs with output resolution. So $0.07 for max is the price of a one megapixel image, and the same image at 4 megapixels costs four times that.

For klein they publish the formula themselves: the first megapixel costs a flat rate and every additional megapixel adds to it. klein 4B is $0.014 plus $0.001 per megapixel, and their own worked example is a 2MP image landing at $0.015. klein 9B has the same shape at $0.015 plus $0.002. The number we put in the category table is that published floor, so if you generate anything larger than a megapixel your invoice will sit above our table.

A contradiction inside their own documentation

For flex, the family overview page writes $0.06 per megapixel in its "which model should I choose" table, and the pricing page in the same documentation set writes "from $0.05". Those are not the same number and we have not picked one. Both pages were read today and both are linked at the top of this page. If you are budgeting for flex, check their pricing calculator yourself and do not lean on either figure.

Which member to take

If you do commercial work and want to run locally, you have exactly one option: klein 4B. It runs on an ordinary consumer graphics card with about 13GB of memory, uses four inference steps and returns in under a second. It was built for high volume and real-time work, and its leaderboard score says exactly that: it is low.

If you stay on the API, pro is the balance point and costs half what max costs. One catch that rarely gets written down: pro is the only member where editing an image costs more than generating one, $0.045 against $0.03. For every other member those two numbers are identical.

Take max when you need web search. It is the only member of the family that can search the web while generating, and that matters for an image that has to reference something real and current. flex, in their own words, is specialised for typography, and its steps and guidance are adjustable.

If you want to train on top of it

Take the Base variants rather than the ordinary ones. Black Forest Labs publishes two separate weight families for klein: the distilled version that is fast, and the Base version that is undistilled and has higher output diversity. They write that Base is meant for fine-tuning, LoRA training and research, and that it is not offered on the public API. The licences carry over: 4B Apache, 9B non-commercial.

Two things that waste people time

First, the numbers. FLUX 3 is newer than FLUX.2 but it is not its successor: FLUX 3 is the video line and is priced per second, while FLUX.2 is the image line. If you want images, the bigger number is not the thing you want.

Second, endpoints. For pro and klein 9B there are two endpoints each: one with a preview suffix where the latest improvements land, and one without, which is a fixed snapshot that does not change. If your output has to stay identical between runs, call the one without the suffix. It is a distinction nobody notices until the day their output shifts for no visible reason.

From Iran

We do not know, and this time the reason differs from Google and OpenAI. Those two publish supported-country lists, and we wrote that Iran is absent from them by reading those lists. Black Forest Labs has no such document: we read its full documentation index today and it contains neither a terms page nor a country list, and the terms URL on the main site returns 404. So the Iran cell for this family is empty, and empty here means there was nothing to read, not that we failed to look.