Main area
Agency
ki-modelle

FLUX.2

Redaktion ·

FLUX.2 — why an image model from Freiburg matters to you

Until recently, generating images meant choosing between two camps. Proprietary APIs like Midjourney, OpenAI’s image endpoints or Google Imagen — black boxes that bill per image and run on someone else’s servers. Or open models like SDXL, which you can self-host but which long lagged behind on image quality and prompt adherence.

FLUX.2 is the attempt to merge both worlds: a model that approaches the proprietary top tier on image quality — and that you can download in several variants, run locally, and in part use commercially for free. It is built by Black Forest Labs (BFL) in Freiburg, currently the most prominent German foundation-model company.

This article explains what FLUX.2 is technically, how the variants differ, what “open-weight” actually means here, and how the model compares to SDXL, Imagen and Google’s Nano Banana. By the end you can judge for yourself whether FLUX.2 fits your workflow — and which variant.

Core mechanics: What FLUX.2 actually is

FLUX.2 is a text-to-image model: you enter a text prompt (“a law office in evening light, photorealistic”) and the model produces an image. It also handles image editing — modifying existing images via instructions — and multi-reference editing, where several reference images serve as input at once (e.g. take the face from image A, the style from image B).

Rectified flow transformer — the architecture in one sentence

Technically FLUX.2 is a rectified flow transformer. This is an evolution of the diffusion idea: instead of “carving” an image out of noise over dozens of steps, a rectified-flow model learns a straighter path from noise to finished image. The practical effect is that fewer compute steps are needed — which directly raises generation speed. That very lever is what makes the small FLUX.2 variants so fast.

The people behind it matter: BFL was founded by Robin Rombach and Andreas Blattmann — the same researchers who did key work on Stable Diffusion at Stability AI. FLUX is therefore not just another vendor but the direct continuation of the lineage that SDXL also came from.

Parameters, resolution, speed

The FLUX.2 family spans a wide range. The largest open model, FLUX.2 [dev], has 32 billion parameters; the smallest open variant, FLUX.2 [klein], gets by with 4 billion. Per Black Forest Labs, the models generate images at up to 4 megapixels and support reference editing with up to ten input images on the pro variant. For the small [klein] variant, BFL promises sub-second inference on capable hardware — a magnitude that turns local image generation from a batch job into an interactive process.

Pitfall: “open source” is not the same as “free for commercial use”

The most common misunderstanding about FLUX.2 concerns licensing. “Open-weight” only means the model weights are publicly downloadable — not automatically that you may use them for anything.

BFL runs a tiered licensing strategy. Only some models sit under the permissive Apache 2.0 license (unrestricted commercial use). Other open variants run under the FLUX non-commercial license — you may download and try them, but commercial products require a paid license or going through the BFL API. The proprietary top variants are API-only anyway.

The variants in detail

FLUX.2 is not a single file but a family. The logic: BFL releases the weights as a distribution channel (reach, community, integrations) and earns money on the proprietary top models and commercial contracts.

FLUX.2 [klein] — the fast, small model

FLUX.2 [klein] is the entry variant with 4 billion parameters under Apache 2.0 — so commercially free. It is built for speed and low hardware barriers: per BFL it runs from around 13 GB of VRAM, i.e. on a single consumer card like an RTX 4060 Ti 16 GB. Quantized versions (FP8, NVFP4) lower the memory footprint further, making 8–12 GB cards realistic too. There is also a larger 9B sub-variant, but it sits under the non-commercial license.

FLUX.2 [dev] — the large open model

FLUX.2 [dev], at 32 billion parameters, is the most capable open FLUX model. BFL positions it as the best open-weight image model for text-to-image as well as single- and multi-reference editing. It needs correspondingly more hardware and is not meant for every consumer GPU. Via partners like Cloudflare Workers AI, [dev] also runs directly at the edge, so you do not have to operate your own GPU infrastructure.

FLUX.2 [pro] and [flex] — the proprietary API models

The pro and flex variants are not open — they are available only through the BFL API. [pro] targets maximum image quality that, per BFL, keeps up with the best closed models while generating faster and cheaper. [flex] gives you more control over steps and guidance scale and is tuned for text rendering and fine detail.

| Variant | Parameters | License | Availability | Typical use | |---|---|---|---|---| | FLUX.2 [klein] 4B | 4B | Apache 2.0 | Open weight | Local, fast, commercially free | | FLUX.2 [klein] 9B | 9B | FLUX non-commercial | Open weight | Local, higher quality, non-commercial | | FLUX.2 [dev] | 32B | partly non-commercial | Open weight | Best open model, cloud/on-prem | | FLUX.2 [pro] | n/a | proprietary | API only | Top quality, production | | FLUX.2 [flex] | n/a | proprietary | API only | Text rendering, control |

Exact license terms change; always check the current repo before use.

Market context: SDXL, Imagen, Nano Banana

FLUX.2 is best understood through comparison. Three reference points help.

Against SDXL — the open predecessor standard

SDXL (Stable Diffusion XL) was for years the default open model for local image generation — a huge community, countless fine-tunes and LoRAs. FLUX comes from the same research lineage (Rombach/Blattmann) but aims considerably higher: better prompt adherence, cleaner anatomy, and above all usable text rendering inside the image — a classic SDXL weakness. If you set up an open model from scratch today, you pick FLUX over SDXL for most use cases. SDXL stays relevant where the massive LoRA library and very modest hardware requirements matter.

Against Imagen — Google’s proprietary model

Google Imagen is a closed, high-quality text-to-image model that runs only through Google’s cloud APIs. The decisive difference is not primarily image quality but the operating model: you cannot download Imagen, cannot run it locally, and cannot use it without token costs. FLUX.2’s open variants give you the option to solve the same task without per-image billing and without data leaving for a cloud provider — relevant anywhere privacy or unit costs matter.

Against Nano Banana — Google’s editing model

“Nano Banana” is the nickname for Google’s editing-strong image model (located in the Gemini family), which became known mainly for consistent editing and character consistency. Here FLUX.2 competes directly via its multi-reference editing. The trade-off is the same as with Imagen: Nano Banana is an API product, whereas you can self-host FLUX.2 [dev] and [klein]. Want maximum editing quality without hosting effort, look at Google; need control, local processing or an Apache 2.0 license, look at FLUX.

For more background on the [klein] release, see the news on FLUX.2 [klein]. If you want to run models locally in general, it is worth looking at Ollama for the LLM side and Hugging Face as the source for the weights.

Practice: Which variant for which case

  • You want to generate locally, fast and commercially free (e.g. variants for agency mockups): FLUX.2 [klein] 4B under Apache 2.0. Runs from ~13 GB VRAM, sub-second inference.
  • You need the best open quality and have hardware or cloud budget: FLUX.2 [dev] (32B), locally on a strong GPU or via Cloudflare Workers AI at the edge.
  • You want to operate nothing yourself and pay per image: FLUX.2 [pro] or [flex] via the BFL API.
  • You need a huge fine-tune/LoRA library on weak hardware: SDXL can still be the more pragmatic choice here.

FAQ

Is FLUX.2 really open source?
Partly. "Open-weight" is more precise: the weights of several variants are freely downloadable. But only some (the 4B [klein] variant) sit under the commercially free Apache 2.0 license. Other open variants run under the FLUX non-commercial license, and the proprietary top models are API-only.
What hardware do I need for FLUX.2?
For FLUX.2 [klein] 4B, BFL says around 13 GB of VRAM is enough — a consumer card like an RTX 4060 Ti 16 GB. With quantized versions (FP8/NVFP4), 8–12 GB cards become realistic too. FLUX.2 [dev] at 32B needs considerably more and is meant for strong GPUs or cloud inference.
Who is behind FLUX.2?
Black Forest Labs (BFL) in Freiburg, founded by Robin Rombach and Andreas Blattmann — the same researchers who did key work on Stable Diffusion. BFL is regarded as the most prominent German foundation-model company in the image space and closed a Series B of $300M at a $3.25B valuation in December 2025.
What sets FLUX.2 apart from SDXL?
Both come from the same research lineage, but FLUX.2 is newer and stronger: better prompt adherence, cleaner anatomy, and above all usable text rendering inside the image — an SDXL weakness. SDXL keeps an edge on very modest hardware and with the huge existing LoRA library.
Can I use FLUX.2 for commercial projects?
Yes, but only with the right variant. FLUX.2 [klein] 4B under Apache 2.0 is commercially free. The 9B variant and parts of [dev] fall under the FLUX non-commercial license — commercial use then runs through the paid BFL license or API.

Conclusion

FLUX.2 is currently the most serious attempt to bring open image models up to the proprietary quality tier — with a clearly tiered strategy: fast small models (Apache 2.0) for reach, a large open model for maximum open quality, proprietary API models as the revenue base. For practice this means: check the license of the specific variant first, then the hardware. If you want to work locally and commercially free, start with FLUX.2 [klein] 4B. If you need maximum quality without running anything yourself, go via API or edge.

As a side note, FLUX.2 is a rare data point: a German company playing at the top tier in a foundation-model domain — and doing so by specializing in images rather than chasing ever-larger language models.

Themenuebersicht

KI-ModelleOpen-Weight