FLUX.2 Explained: Dev, Pro, Flex, Max, Klein and Which to Use

The short answer

FLUX.2 is Black Forest Labs' second-generation image model family, launched on 25 November 2025. It comes in five tiers, and which one you can use depends less on quality than on licence:

• [pro], [flex] and [max] are API-only. You pay per image (from $0.03, $0.05 and $0.07 at the time of writing) and never see the weights.
• [dev] is a 32B open-weight model under the FLUX Non-Commercial License. Free to download, not free to run as a business.
• [klein], added on 15 January 2026, is the small, fast branch. klein 4B is Apache 2.0, so you can self-host it commercially. klein 9B is non-commercial, like [dev].

If you want one rule: use the API tiers if you just need images, klein 4B if you need to run the model yourself for money, and [dev] or klein 9B only for research and personal projects. The rest of this page explains why, as of October 2026.

From FLUX.1 to FLUX.2: what changed

Black Forest Labs (BFL) makes the FLUX models. Its first family, FLUX.1, arrived on 1 August 2024 in three tiers: [pro] through the API, [dev] as open weights under a non-commercial licence, and [schnell] as a fast Apache 2.0 model. All public FLUX.1 models were 12B parameters.

FLUX.2 keeps the tier idea. The main changes are below; on top of them, FLUX.2 outputs up to 4 megapixels, and the open [dev] model reads prompts with a 24B Mistral-3 vision-language model as its text encoder.

"Rectified flow" is the training method all FLUX models describe on their model cards. Instead of learning to remove noise step by step, the model learns a path from random noise to an image and is pushed to make that path as straight as possible, so fewer steps are needed. That is part of why the small klein models can render in a handful of steps.

The five FLUX.2 tiers at a glance

A few details that matter in practice:

• Reference limits differ by tier. The [pro] API takes up to 8 reference images. BFL's docs recommend a maximum of 6 for [dev]. klein accepts up to 4 through the API.
• [max] is the only tier with web grounding. BFL describes it as generation grounded by a web search, useful when a prompt refers to something recent.
• [dev] has no per-image price from BFL. It is meant to be run on your own hardware or through a licensed host.

FLUX.2 klein in depth

klein is not a separate product line. It is FLUX.2 shrunk down, released in two sizes, each in two forms.

4B against 9B. The 4B model is the one most people should look at first: it is Apache 2.0 and needs about 13 GB of GPU memory. The 9B model pairs a 9B image model with an 8B Qwen3 text embedder; BFL claims it matches models five times its size. It is under the non-commercial licence.

Distilled against Base. Each size ships twice:

Distillation trains a "student" model to reach in a few steps what the full model reaches in dozens. BFL says the cost is output diversity: the distilled models give less varied results for the same prompt, while the Base models keep higher diversity and are the ones to fine-tune.

One model for three jobs. klein does text-to-image, single-image editing and multi-reference editing in the same weights. You do not swap models to go from "make this" to "change this".

Faster builds. BFL also published FP8 and NVFP4 versions with NVIDIA, which it says run up to 1.6 and 2.7 times faster on RTX 5080 and 5090 cards. On 4 June 2026 BFL announced klein running on-device on ASUS ProArt laptops.

Licences in plain words

This is the part most articles get wrong, so here it is without hedging.

Three points worth knowing:

1. Open weights are not open source. [dev] is downloadable, but its licence does not allow commercial use of the model. BFL sells that separately.
2. Outputs and models are treated differently. BFL's [dev] model cards say the images you make may be used commercially; it is running the model as a service that needs the paid licence.
3. Fine-tunes inherit the licence. A LoRA trained on [dev] or klein 9B carries the same non-commercial limits as its base model. If you want a commercial fine-tune, train on klein Base 4B.

The non-commercial licence also asks you to run content filters or manual review. BFL's own model cards describe filtered training data (with known abuse material removed in work with the Internet Watch Foundation), input and output filters in its code, and C2PA provenance metadata on every API image. On the API, a safety_tolerance setting runs from 0 (strictest) to 5, with 2 as the default, and a request blocked at the input stage is not charged. Why mainstream tools draw these lines is covered in why AI image generators block adult content.

Kontext, and how FLUX.2 folded editing in

FLUX.1 Kontext, released on 29 May 2025, was BFL's answer to "change this picture, keep everything else". You gave it an image and an instruction such as "make it winter" or "put her in a red coat", and it returned the edit. Its paper describes the trick: the reference image is turned into tokens and placed in the same sequence as the image being generated, so the model looks at the reference while it draws. That gave it strong character preservation across scenes and little drift over repeated edits. Kontext [pro] cost $0.04 per image and [max] $0.08; an open Kontext [dev] followed under the non-commercial licence.

FLUX.2 makes that the default. Every tier accepts reference images, and [pro] takes up to 8 at once. In practice that means one model can place the same character in a new scene, combine a person, a product and a background from three pictures, or apply a precise edit. If you are trying to keep a character recognisable across many images, our guide to consistent AI characters compares this reference-image approach with the other three common ones.

Which FLUX.2 should you use?

If FLUX is only one of the models you are weighing, our comparison of the best AI image generators puts it next to GPT Image, Nano Banana, Midjourney, Seedream and the open-weight alternatives.

What comes next: FLUX 3

BFL announced FLUX 3 on 23 July 2026. It is one model family trained jointly on image, video and audio rather than three separate models. Video went into early access first, with clips of up to 20 seconds including sound; image access was promised "in the following weeks". BFL's pricing page already lists FLUX 3 images at $0.041 (768 by 768) up to $0.607 (about 16 megapixels). An open-weight "FLUX 3 Dev" has been announced but, as of 14 October 2026, not released, and its licence was not stated.

Until it ships, FLUX.2 is the current open-weight FLUX, and klein 4B remains the only FLUX.2 model you can run commercially without a separate agreement.

How we checked

We checked Black Forest Labs' own announcements, documentation, prices and licences on 14 October 2026. We did not benchmark the models, so nothing here ranks image quality beyond what BFL itself claims.

The easier way: SecretCrush AI

Running FLUX.2 yourself means picking a tier, reading a licence, finding a GPU and tuning prompts before you see a single good picture. If what you actually want is great images of a character, SecretCrush AI does the whole job for you.

• Fast. High-quality character images in about 10 to 15 seconds, and videos in about 40 seconds per clip.
• No set-up. No installs, no GPU, no workflows, no prompts to tune.
• Consistent. The same character keeps the same look from one picture to the next.
• Alive. You chat with your character in immersive, realistic roleplay and ask for a picture or a video right inside the conversation.

A free account gets 10 messages a day, every day. Premium gives you unlimited messages with advanced memory and 500 bonus coins every month (600 on the yearly plan), from $8.99 a month on the yearly plan or $16.99 month to month. An image costs 6 coins, and coin packs ($10 for 300 coins) are open to free accounts too.