Particle.news
Download on the App Store

Black Forest Labs Debuts FLUX.2, Advanced Image Models With Open Weights and Managed APIs

NVIDIA FP8 optimizations cut VRAM needs by roughly 40%.

Overview

  • The FLUX.2 family spans pro, flex, and a 32B-parameter dev model, with open weights for FLUX.2 [dev] on Hugging Face and hosted access via partners such as TogetherAI, Replicate, Runware, Verda, Cloudflare, DeepInfra, and FAL.
  • Cloudflare Workers AI now serves FLUX.2 [dev] with multipart form-data for multi-image inputs (up to four 512×512 images) and output resolution up to 4 megapixels.
  • Core capabilities include multi-reference consistency (up to 10 images per BFL), improved prompt adherence, cleaner typography, direct pose control, and unified generation-plus-editing up to 4MP.
  • The full 32B checkpoint is resource intensive—about 90GB VRAM to load or ~64GB in lowVRAM—offset in part by FP8 quantization and ComfyUI weight streaming for RTX PCs, with performance tradeoffs.
  • BFL positions pro and flex for production APIs and control, while the dev weights are reported as non‑commercial, with a size‑distilled FLUX.2 [klein] model listed as coming soon.