FLUX.2: Frontier Visual Intelligence (bfl.ai)

Black Forest Labs today launched FLUX.2, a family of image-generation and editing models aimed at real-world production workflows rather than demos. FLUX.2 emphasizes multi-reference consistency (up to 10 inputs), reliable typography and logo handling, structured prompt following, and image editing up to 4 megapixels. The suite includes managed, production-ready APIs (FLUX.2 [pro]), a controllable developer tier with step/guidance parameters (FLUX.2 [flex]), a 32B open-weight checkpoint for researchers and hobbyists (FLUX.2 [dev] on Hugging Face and multiple APIs), and a forthcoming compact Apache‑2.0 distilled variant (FLUX.2 [klein]). The company also published a new FLUX.2 VAE under Apache 2.0 to improve the latent trade-offs between learnability, quality, and compression. Technically, FLUX.2 combines a Mistral‑3 24B vision‑language model with a rectified flow transformer in a latent flow‑matching architecture, retraining the latent space from scratch to boost realism, lighting consistency, spatial logic, and text rendering. FLUX.2 [dev] claims state‑of‑the‑art open‑weight performance across text‑to‑image and multi‑reference editing and can run locally on a single RTX 4090 with an fp8 reference implementation (collab with NVIDIA/ComfyUI). For the AI/ML community this means production-grade, inspectable image models that lower costs, enable local experimentation, and push open innovation — while offering managed endpoints for teams needing scale and reliability.