Black Forest Labs releases FLUX.2: 32B open-weight image model with 4MP editing and 10-image multi-reference support
Black Forest Labs has released FLUX.2, a family of image generation models including a 32B parameter open-weight variant. The models support editing at up to 4 megapixel resolution and can reference up to 10 images simultaneously for character and style consistency.
Black Forest Labs Releases FLUX.2 Image Generation Models
Black Forest Labs has released FLUX.2, a family of image generation and editing models that includes a 32-billion parameter open-weight variant capable of editing images at up to 4 megapixel resolution.
Model Specifications
The FLUX.2 family includes four variants:
FLUX.2 [pro]: Closed-source production model available via API at BFL Playground and launch partners. The company claims state-of-the-art image quality matching competing closed models while generating images faster and at lower cost. Pricing not yet disclosed.
FLUX.2 [flex]: API-accessible model with adjustable parameters including step count and guidance scale for controlling quality-speed tradeoffs. Available now through BFL API and partners. Pricing not yet disclosed.
FLUX.2 [dev]: 32B parameter open-weight model available on Hugging Face under a commercial license. This is the most powerful variant released with open weights. The model runs on consumer GPUs including GeForce RTX cards using an optimized fp8 implementation developed with NVIDIA and ComfyUI. Also available via API through FAL, Replicate, Runware, Verda, Together AI, Cloudflare, and DeepInfra.
FLUX.2 [klein]: Upcoming Apache 2.0 licensed model, distilled from the FLUX.2 base. Beta access available by sign-up.
Technical Capabilities
All FLUX.2 variants combine text-to-image generation and image editing in a single architecture. Key capabilities according to Black Forest Labs:
- Multi-reference support: Process up to 10 reference images simultaneously for character, product, and style consistency
- Resolution: Generate and edit images up to 4 megapixels
- Text rendering: Complex typography, infographics, and UI mockups with legible fine text
- Flexible aspect ratios for both input and output
Architecture
FLUX.2 uses a latent flow matching architecture that couples the Mistral-3 24B parameter vision-language model with a rectified flow transformer. The company built a new variational autoencoder (FLUX.2-VAE) from scratch to optimize the tradeoff between learnability, quality, and compression rate. The VAE is released under Apache 2.0 license on Hugging Face.
According to Black Forest Labs, the vision-language model component provides real-world knowledge and contextual understanding, while the transformer handles spatial relationships, material properties, and compositional logic.
Performance Claims
Black Forest Labs states that FLUX.2 [dev] "sets a new standard" for open-weight image models, outperforming all open-weight alternatives across text-to-image generation, single-reference editing, and multi-reference editing benchmarks. Specific benchmark scores were not provided.
The company claims FLUX.2 [flex] allows quality-speed tradeoffs through a "steps" parameter, with examples showing 6, 20, and 50-step generation options affecting typography accuracy and image detail.
Availability
FLUX.2 [pro], [flex], and [dev] are available now. FLUX.2 [dev] weights can be downloaded from Hugging Face. The FLUX.2-VAE is available under Apache 2.0 license. FLUX.2 [klein] is in beta.
What This Means
FLUX.2 [dev] represents the largest open-weight image generation model released to date at 32B parameters, and the first to combine generation and editing with multi-reference support in a single checkpoint. The 4MP editing capability and 10-image multi-reference feature are significant technical advances, though independent benchmarks will be needed to verify the company's performance claims against competitors like Stable Diffusion 3 and other open models. The Apache 2.0 licensed VAE and upcoming [klein] model suggest a commitment to open-source infrastructure, while the commercial [pro] and [flex] variants follow Black Forest Labs' hybrid business model of pairing open research with commercial offerings.
Related Articles
Nvidia releases Nemotron 3.5 Super VL: open 120B multimodal MoE with 12B active parameters
Nvidia has released Nemotron 3.5 Super VL, an open multimodal Mixture-of-Experts model with 120 billion total and 12 billion active parameters, built for continuously running enterprise agents. Two compact Bolt models for local PCs and edge devices launch alongside it, though the 4 GB variant ships later this month.
Mistral's 209M-parameter Timestral tops GIFT-Eval, TIME and fev-bench time-series leaderboards
Mistral has released Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. Mistral claims the largest model ranks first among pretrained zero-shot models on GIFT-Eval, TIME and fev-bench, and runs on CPUs without a GPU.
OpenBMB's MiniCPM-V 4.7 35B-A3B MoE claims 13+ tokens/s on iPhone 17 Pro Max in 3 GB
OpenBMB has released MiniCPM-V 4.7 35B-A3B, the first mixture-of-experts model in the MiniCPM-V series, with open weights on Hugging Face. The company claims it runs at 13+ tokens/s using 3 GB of memory on an iPhone 17 Pro Max, and that it beats Qwen3.6 35B-A3B on overall score in less inference time in Instruct mode.
Tencent releases Hy-Memory-30B-A3B, a 3B-active model that writes agent memory for Mem0 and MemOS
Tencent Hunyuan released Hy-Memory-30B-A3B, a 30B-parameter mixture-of-experts model with about 3B active parameters, built to run the write path in external memory frameworks. Tencent reports a 55.49 cross-harness average, ahead of Claude Opus 5 (54.62) and DeepSeek V4.1 Flash (53.64) on its own fixed 20% benchmark subsamples.
Comments
Loading...