FLUX 3 Image Klein: Open-Weight Image AI on an 8GB GPU

Black Forest Labs released FLUX 3 Image Klein on October 15, 2026. It is a 5.6-billion-parameter image model with open weights that handles text-to-image, image editing and multi-reference generation from up to ten reference images, outputs at native 4K, and accepts bounding boxes that place an element or confine an edit to one region. The company publishes it in two forms, a base model for fine-tuning and a step-distilled model for speed, and in FP8, INT8 and NVFP4 builds developed with NVIDIA. The INT8 build runs on GeForce RTX cards with as little as 8 GB of video memory through ComfyUI or Diffusers. The licence is free for commercial use below $5 million in annual revenue.

This is the second open-weight release in the FLUX 3 family, after FLUX 3 Action, and the one most marketing teams can run on the machine under the desk. Our FLUX 3 launch post covers the full multimodal model, and our Nano Banana 2.1 versus FLUX 3 Image comparison tests the hosted version on product content. Klein is the small sibling, and smallness is the point: it changes who can run the model and what it costs, not what the model is for.

Everything below is Black Forest Labs' own statement unless marked otherwise, read from the launch post on October 16. The speed figures are NVIDIA's measurements where the post says so. The ranking claims are internal evaluations and are treated as such. We have not run the model and make no quality claim of our own.

The launch post describes one model in several packages. The base model is for fine-tuning and runs 50 denoising steps. The step-distilled model is for low-latency generation and runs 16. The post says Klein shares core capabilities with the full FLUX 3 Image model: native 4K, pixel-perfect editing and bounding boxes. The quantised builds were developed with NVIDIA support, the model was trained on NVIDIA GB200 NVL72 systems, and NVIDIA supplies serving through TensorRT-LLM and fine-tuning recipes through NeMo AutoModel. Day-one support is listed for ComfyUI, Hugging Face Diffusers, TensorRT-LLM, SGLang Diffusion and vLLM-Omni, with LoRA training through the Ostris AI Toolkit.

What it is not: a video model, a replacement for the full FLUX 3 Image model on the API, or a phone model; the post's on-device example is laptops. It is a small image model that trades some capacity for the ability to run where the previous generation could not. Our guide to local image generation with ComfyUI covers the toolchain this drops into; Klein is new since that guide was written.

Three controls matter for commercial imagery and Klein has all three. Multi-reference generation accepts up to ten reference images, so a product, a model, a background and a brand palette can each be supplied as a picture rather than described in words. Native 4K output means a hero image does not need a separate upscaling pass. And bounding boxes let you say where an element goes or fence an edit inside a region, which the post describes as placing elements or confining edits to a defined area.

Why the box matters most: product imagery lives or dies on the product staying exactly as photographed. An edit that changes the background and also softens the logo is a rejected asset. A bounding box that confines the edit to the background, combined with what the API calls pixel restoration, which the post says preserves unedited regions exactly, is the mechanism that makes a batch of variants reviewable rather than re-shot. The example is a hundred seasonal backgrounds behind one bottle. The failure case is the box drawn too tight: a shadow or reflection the product casts outside the region will not update, and the composite reads as pasted. Draw the box around the product and its shadow, not the product alone, and check every take at 100% zoom.

Black Forest Labs gives the generator's weight sizes directly: 11.2 GB in BF16 and 5.6 GB in FP8. It gives no file size for the INT8 or NVFP4 builds, but states that the INT8 build runs on GeForce RTX GPUs with as little as 8 GB of VRAM through ComfyUI or Diffusers. Those figures are the generator alone; the text encoder and working memory sit on top, a likely reason the stated INT8 minimum is 8 GB; the post gives no breakdown. The parameter counts the post uses for comparison, including Klein's own 5.6B, exclude text encoders by the vendor's stated convention.

The practical reading for a marketing team: a workstation or gaming PC bought since 2023 with an RTX 4070 or better runs the distilled model locally at usable speed, and an 8 GB card runs it at all. The failure case is reading "8 GB" as the comfortable configuration. It is the floor, on the INT8 build, with the quality cost quantisation carries; Black Forest Labs publishes no quality comparison between its FP8, INT8 and NVFP4 builds.

The post gives two kinds of speed figure and they come from different hands. The headline pair, about four seconds for a one-megapixel image on an RTX 5090 and under nine seconds on an RTX 4070 with the distilled INT8 model, is Black Forest Labs' summary. The detailed tables are attributed to NVIDIA, with the RTX results taken in ComfyUI: for a 1024-square text-to-image at 16 steps the distilled model runs from 0.55 seconds on a GB300 to 8.9 seconds on an RTX 4070, and the base model at 50 steps from 1.7 seconds on a GB300 or DGX Station to 43.8 seconds on a DGX Spark. Serving a LoRA adds about 5% on a GB300.

For the hosted API the post reports its own median end-to-end latency at one megapixel: 4.54 seconds for text-to-image and 7.18 seconds for editing, measured on Black Forest Labs' servers. It then compares against competitors timed from the client, including polling and network time, which is not the same measurement; the post itself notes the difference. Read the competitor latencies as illustrative and the Klein figures as the vendor's own.

The weights ship under the FLUX Kommunity License. The launch post summarises it in three parts: free commercial use for individuals and companies with annual revenue under $5 million; research and non-commercial use; and, for larger organisations, commercial deployment through self-hosted licensing or the hosted API. We read the post's summary, not the licence file itself, so the conditions attached to the threshold, how revenue is measured, what happens when you cross it, and what attribution is owed, are questions for the licence text. Our open-weight licence audit compares revenue-capped licences across vendors. Meta's Muse Spark Open, released the same week, uses a $20 million threshold, but it applies only to model-as-a-service providers, so the two lines measure different things.

Why the line matters for agencies in particular: $5 million in annual revenue is a threshold a mid-sized agency, and many of its clients, can exceed. The example is a boutique studio under the cap that runs Klein locally for a client's campaign; the studio is inside the licence as the post describes it, and the client receives finished images. The failure case is the same studio deploying the weights inside a client's own pipeline, where the client's revenue may be the one that counts. Read the licence file before the second arrangement.

The hosted model is called flux-3-image-klein and is the distilled version, with two additions the post lists for the API: prompt upsampling, which expands a short prompt before generation, and pixel restoration, which the post says preserves unedited regions exactly. The price line in the launch post reads, in full, "Pricing starts at $0.009 per 1K image". Black Forest Labs' pricing table, read October 16, defines the unit: per image by output resolution, $0.008 at 768 × 768, $0.009 at 1K, $0.011 at 1.5K, $0.014 at 2K and $0.036 at 4K, with text-to-image and editing priced the same.

Why it matters for budgeting: at 1K resolution the model cost of a variant-heavy campaign is under a cent an image and negligible next to the review time. The example is a retailer generating twenty background variants for each of 500 products, 10,000 images: about $90 at 1K, or $360 at 4K. The failure case is budgeting the whole campaign at the 1K rate and discovering that 4K hero images cost four times as much. Price the resolutions you will actually order before quoting a client.

Black Forest Labs says Klein ranks first among open-weight models in text-to-image, multi-reference generation and image editing, ahead of larger open models, and rates above the closed Ideogram V4.5 and its own FLUX.2 [pro] in the same comparison. Against its predecessor, FLUX.2 [klein] 9B, it reports gains of 91 Elo in text-to-image and 70 Elo in editing with 38% fewer parameters. The post labels these as internal evaluations, names the competitor parameter counts as each publisher reports them, and notes that Krea 2 is excluded from the editing comparison because it does not edit.

Two labels belong on all of it. It is the vendor's own evaluation, run by the vendor on prompts the vendor chose, which is the weakest tier of evidence a buyer can act on. And an Elo gain is relative: 91 points over the previous small model says the small model got better, not that it matches the full FLUX 3 Image. The one external check the post cites is on safety, not quality: an evaluation by Cinder that the post summarises as finding "substantially fewer vulnerabilities than other popular open weights models", alongside dataset filtering and internal red-teaming. For a brand deploying the model in a customer-facing tool that is the more useful sentence, and it is still a vendor summarising its own partner.

No "best image model" verdict here; we have not generated a single image with it and the quality claims are the vendor's. What the hardware, licence and price statements do support is a set of fits.

If your team produces product or campaign imagery and is weighing a locally run open-weight model against the hosted services it already pays for, the licence reading, the hardware fit and the side-by-side quality test on your own products are work our content engine team does.