Black Forest Labs drops FLUX.2 with 10-image referencing and open 32B weights

FLUX.2 is a new family of visual generation and editing models that handles up to 10 reference images at once and offers an open-weight 32B checkpoint.

Built for practical creative workflows rather than quick party tricks, it unifies text-to-image synthesis and multi-image editing in a single model while supporting up to 4MP resolutions.

What it does

The marquee feature is multi-reference consistency. You can feed the model up to 10 reference images simultaneously, letting you preserve character facial features, specific product details, or exact visual styles across varied compositions without complex external setups.

Text rendering also takes a notable leap forward. FLUX.2 handles fine typography cleanly, producing legible text for UI mockups, infographics, and poster designs where earlier diffusion models generated gibberish.

The lineup splits into four models: the flagship FLUX.2 [pro] API for top visual fidelity, FLUX.2 [flex] which lets users tune generation steps to trade latency against detail, the 32B open-weight FLUX.2 [dev], and a smaller distilled model named FLUX.2 [klein] coming soon under Apache 2.0.

Under the hood

FLUX.2 pairs a 24B Mistral-3 vision-language model with a rectified flow transformer using a latent flow matching architecture.

Mistral-3 provides deep semantic comprehension and real-world spatial logic, while the flow transformer executes the fine-grained visual composition, lighting, and textures.

The team also completely retrained its latent representation, releasing the standalone FLUX.2 - VAE under an Apache 2.0 license on Hugging Face to optimize the tradeoff between compression and fidelity.

Requirements and availability

The 32B FLUX.2 [dev] weights are live on Hugging Face with reference inference code. Running a model of this size locally requires serious hardware, but a native fp8 implementation developed with NVIDIA and ComfyUI allows it to run on consumer GeForce RTX GPUs.

FLUX.2 [dev] is open for non-commercial exploration, with commercial licenses offered separately on the BFL website.

For those who prefer managed infrastructure, the pro and flex variants are available on BFL Playground as well as through API providers like FAL, Replicate, Runware, and TogetherAI.

Why it matters

For designers, agency workflows, and game developers, native 10-image referencing makes character turnaround sheets and commercial product staging drastically faster without training dedicated LoRAs.

In our view, Black Forest Labs continues to hold the line for open creative tooling. By putting a competitive 32B model into open weights alongside an fp8 pipeline, it prevents high-fidelity generation from becoming an exclusive feature of closed proprietary ecosystems. If you have the GPU headroom, it is well worth testing locally right away.