FLUX 3 Explained: Release Status and Local Setup

FLUX 3 is a multimodal generative model family developed by Black Forest Labs that spans image generation, image editing, video synthesis with synchronized audio, and robotic world action modeling.

Black Forest Labs began rolling out the FLUX 3 architecture in mid-2026, dividing capabilities between paid API endpoints and open-weight models such as FLUX 3 Image Klein on Hugging Face.

The lineup splits into the hosted FLUX 3 Image model, the 5.6-billion-parameter open-weights FLUX 3 Image Klein, and separate Video and Action models.

What FLUX 3 Is and Which Parts Are Out Today

FLUX 3 is a multimodal flow model family trained jointly on images, video, audio, and action prediction by Black Forest Labs. Black Forest Labs introduced the FLUX 3 architecture on July 23, 2026. As of October 2026, four FLUX 3 models are out, and FLUX 3 Dev has no release date.

The flagship image generation tool, FLUX 3 Image, launched in early October 2026 as a managed API endpoint. Shortly after, Black Forest Labs released FLUX 3 Image Klein on October 15, 2026, providing downloadable weights for local inference and fine-tuning. FLUX 3 Video came earlier. Black Forest Labs published its Video announcement on August 4, 2026, and the model is generally available on its API. FLUX 3 Action, a 7B open-weights world action model, followed on September 23, 2026. VentureBeat reported at launch that FLUX 3 started in limited release.

Black Forest Labs has announced FLUX 3 Dev as an upcoming open-weights multimodal backbone. However, Black Forest Labs has not announced a release date, parameter size, or licensing terms for FLUX 3 Dev as of October 2026. The video model is covered in FLUX 3 Video explained.

Editing Features and Controls in FLUX 3 Image

FLUX 3 Image is Black Forest Labs' flagship image model, sold through a managed cloud API and, for companies at scale, a commercial weights license from sales. One endpoint handles text-to-image, single-image editing and multi-reference composition, so a single integration covers all three jobs.

FLUX 3 Image accepts up to 10 reference images in one request, according to the FLUX 3 Image model page. Reference inputs are tagged sequentially in prompt syntax as <ref_image_0>, <ref_image_1>, up through <ref_image_9>. Black Forest Labs rejects input reference files larger than 16 Megapixels (MP) with an HTTP 400 error, while images smaller than 256 by 256 pixels trigger an HTTP 422 error.

Bounding-box edits use a 0 to 1000 grid on both axes. According to the bounding box specifications, developers format coordinate boxes as [y_min, x_min, y_max, x_max]. Bounding boxes let developers place objects, recolor designated items, replace assets, or remove elements while the surrounding pixels stay unchanged. An internal prompt upsampler expands brief instructions into descriptive scene captions without overriding user-drawn coordinate boundaries. Generated outputs arrive as signed URLs that expire after one hour. Company licensing and rollout are covered in FLUX 3 for business.

Open Weights in FLUX 3 Image Klein

FLUX 3 Image Klein is a 5.6-billion parameter open-weights image generation model built for efficient local execution and customization. Released on October 15, 2026, the model is detailed in the Black Forest Labs Klein release announcement. That 5.6B figure covers the image generator only; text encoders are not counted.

Black Forest Labs provides two distinct versions of the architecture in the Klein technical documentation. The standard Klein model is a distilled checkpoint running in 16 inference steps with guidance scale 1.0, designed for rapid generation and consumer hardware. The Klein Base model is an undistilled checkpoint running across 50 inference steps with guidance scale 3.5, optimized for Low-Rank Adaptation (LoRA) training, checkpoint fine-tuning, and architectural research. Both variants support multi-reference workflows with up to 10 input images, 0 to 1000 bounding-box coordinates, and resolutions up to 4K.

Black Forest Labs distributes Klein under the FLUX Kommunity License on Hugging Face. The license grants free commercial use for individuals and organizations generating under five million dollars in annual revenue. Entities exceeding five million dollars in annual revenue must license self-hosted weights or use the managed API. In Black Forest Labs' internal Elo evaluations, Klein ranked first among the open models tested in text-to-image, multi-reference generation and editing. The same tests rate it above two closed models, Ideogram V4.5 and FLUX.2 [pro]. Against FLUX.2 [klein] 9B, Black Forest Labs reports an Elo improvement of +91 in text-to-image and +70 in editing while reducing parameters by 38 percent. Earlier FLUX generations are covered in FLUX explained.

Local GPU Hardware Requirements for Klein

Black Forest Labs says the INT8 build of FLUX 3 Image Klein runs on NVIDIA GeForce RTX GPUs with as little as 8 GB of VRAM, through ComfyUI or Diffusers. The generator weights are 11.2 GB in BF16 and 5.6 GB in FP8. Quantized INT8 and NVFP4 formats were developed in collaboration with NVIDIA.

Black Forest Labs reports that the distilled INT8 build renders a 1 MP image in ComfyUI in about 4 seconds on an NVIDIA GeForce RTX 5090. On an RTX 4070, the same image takes under 9 seconds. Across Black Forest Labs' hardware table, distilled 16-step 1024 by 1024 generation ranges from 0.55 seconds on an NVIDIA GB300 to 8.9 seconds on an RTX 4070. The undistilled Klein Base model at 50 steps ranges from 1.7 seconds on GB300 up to 43.8 seconds on DGX Spark. Hugging Face Diffusers is supported from day one, and LoRA training runs on the Klein Base weights.

Black Forest Labs says not to use Klein without its prompt upsampler (black-forest-labs/flux-3-image-klein-pup), because Klein was trained on the captions that model writes and a prompt sent as typed gives noticeably weaker results. The upsampler is a separate model that the Diffusers pipeline does not call, and it adds about 28 GB of GPU memory. Local operators on 8 GB or 16 GB graphics cards must either provide detailed scene captions manually or run the upsampler separately. ComfyUI node setup is covered in ComfyUI explained.

API Integration and Output Tiers

FLUX 3 Image runs at the /v1/flux-3-image endpoint and Klein at /v1/flux-3-image-klein, according to the FLUX 3 Image API docs. Both take the same request fields and return a polling task that ends in a signed image URL.

Billing is pay-as-you-go in credits, and one credit equals $0.01, per the Black Forest Labs pricing page. Image generation charges depend on output resolution tiers, which maintain equal pixel areas across different aspect ratios. Resolution tiers include 768sq, 1k (the default), 1.5k, 2k, and native 4k, with aspect ratios spanning 21:9, 16:9, 4:3, 1:1, and auto.

At the standard 1K resolution tier (approximately 1 MP), FLUX 3 Image costs $0.048 per image, while FLUX 3 Image Klein costs $0.009 per image. At 4K resolution (reaching 5456 by 3072 pixels or 16.8 MP), FLUX 3 Image costs $0.607 per render compared to $0.036 on Klein. Generating 1,000 images at 1K costs $48 on FLUX 3 Image and $9 on Klein, while 1,000 images at 4K totals $607 on FLUX 3 Image versus $36 on Klein. Black Forest Labs reports median API latency for Klein at 1 MP of 4.54 seconds for text-to-image and 7.18 seconds for editing. Every resolution tier is priced in FLUX 3 Image pricing.

Selection Guide for FLUX 3 Deployments

Each FLUX 3 model maps to one kind of job, and the main trade-off is per-image API fees against buying and running a GPU. Klein is a poor fit if you can neither run its 28 GB prompt upsampler nor write full scene captions yourself. A FLUX 3 Dev release with published weights and license terms would change this split.

Match the job to the model:

• Choose FLUX 3 Image via managed API for layout-controlled edits at scale, native 4K outputs up to 16.8 MP, and multi-reference editing.
• Choose FLUX 3 Image Klein for self-hosting or custom LoRA fine-tuning; the INT8 generator fits an 8 GB GPU, but the upsampler needs far more. Klein also suits high-volume API generation at $0.009 per 1K image. Free self-hosted commercial use applies only under $5M in annual revenue.
• Choose FLUX 3 Video for moving media sequences requiring synchronized audio and clip lengths up to 20 seconds on BFL's API.
• Choose FLUX 3 Action for physical simulation modeling, robotics control pipelines, or game engine integrations requiring an open-weights 7B world action model.
• Choose FLUX.2 endpoints if existing production systems require pinned snapshot models like flux-2-pro or flux-2-klein-9b for repeatable generation, as outlined in the FLUX.2 overview.

Current Limitations and Unreleased Components

Several parts of the FLUX 3 family remain restricted or unreleased as of October 2026. The most prominent gap is FLUX 3 Dev, which Black Forest Labs has announced as an open multimodal backbone without publishing release schedules, parameter counts, or licensing structures.

In addition, the flagship FLUX 3 Image model has no public open weights. Organizations wanting to self-host the flagship system must negotiate a custom commercial enterprise license directly with Black Forest Labs sales representatives. FLUX 3 Video has no public weights either. Access runs through BFL's API. The FLUX 3 Image vs FLUX.2 comparison covers what changed between generations, and FLUX alternatives covers competing generators.

Finally, all published benchmark rankings and Elo scores stem directly from Black Forest Labs' internal evaluations rather than independent third-party leaderboards. Production teams should run side-by-side prompt batches to confirm visual fidelity and prompt alignment before migrating software pipelines. To deploy FLUX 3 in an automated pipeline, test a sample batch on the Black Forest Labs API or download the Klein weights and time inference on a local GPU.