FLUX.2

What FLUX.2 is

FLUX.2 is a family of image generation and editing models released by Black Forest Labs (BFL) on 25 November 2025. It generates and edits images at up to 4 megapixels and accepts up to 10 reference images at once, so a single request can combine a written instruction with a stack of supplied pictures.^[1]

The variants

The launch covered several tiers rather than a single model. FLUX.2 [pro] and FLUX.2 [flex] are offered through the BFL Playground, the API and partners. FLUX.2 [dev] is a 32B open-weight model published on Hugging Face, and FLUX.2 [klein] was announced as an upcoming Apache 2.0 model.^[1] A FLUX.2 [max] model was listed from December 2025.^[2]

Openness differs by tier. The FLUX.2 VAE (variational autoencoder), the component that compresses images into a compact representation and decodes them back, was released on Hugging Face under Apache 2.0, while FLUX.2 [dev] requires a separate license for commercial use.^[1]

FLUX.2 [klein] arrived on 15 January 2026 in two sizes: 4B under Apache 2.0 and 9B under the FLUX Non-Commercial License. Both were aimed at generation and editing in under a second.^[3][2]

How it is built

FLUX.2 pairs a Mistral-3 24B vision-language model with a rectified flow transformer.^[1] In broad terms, the vision-language model reads the text prompt and any reference images and turns them into conditioning, while the transformer produces the picture. Rectified flow is a training approach that teaches a model a comparatively straight path from random noise to the finished image; a straighter path generally means fewer sampling steps are needed to reach a usable result.

What the release emphasized

BFL said every FLUX.2 variant offers image editing from text and from multiple references within a single model, rather than splitting generation and editing into separate systems.^[4] The release put particular weight on text rendering — complex typography, infographics, memes and UI mockups — and on improved adherence to structured, multi-part prompts, the kind that specify several subjects, labels or layout requirements at once.^[4]

Where it sits

FLUX.2 follows the FLUX.1 line. BFL described FLUX.1 [dev] as the most popular open image model globally and said FLUX.1 Kontext [pro] powers teams from Adobe to Meta.^[4] The family's position was mixed in outside measurement: in August 2026, Artificial Analysis ranked FLUX.2 15th of 148 image models.^[2] A successor, FLUX 3, followed in July 2026.^[2]