FLUX 3 explained: every model, what it costs and where it runs

Key takeaways

• FLUX 3 is one Black Forest Labs model trained on image, video and audio together. It has shipped in stages: Video on August 4, Image on October 1 and the open-weights Klein on October 15, 2026.

• FLUX 3 Image costs $0.048 per 1K image on BFL's API. Klein costs $0.009 at the same size.

• Klein is the only FLUX 3 image model you can download. Its license lets businesses under $5 million in annual revenue use the images it makes commercially for free.

• FLUX 3 Max is a video model that fal post-trained with Black Forest Labs and serves on its own platform.

• BFL's consumer Playground closed on August 31. The Dev Playground in BFL's dashboard replaced it.

What FLUX 3 is, and why the names get confusing

Black Forest Labs announced FLUX 3 on July 23, 2026, as a single model trained jointly on image, video and audio. It didn't ship all at once. Over three months the company has released it in pieces, and each piece has its own name, price and license. If you last looked during the FLUX.2 era, the lineup has changed a lot.

The short version: FLUX 3 Image is the one for finished stills through an API, Klein is the one to run on your own GPU, and anything with "Video" or "Max" in the name makes clips, not pictures. The sections below take the models one at a time, with prices and hosts for each.

The FLUX 3 lineup at a glance

Prices for the first three are BFL's own API rates, from its release notes and FLUX 3 Video page. There's also FLUX 3 Action, a 7B open-weights robotics model released September 22. It was technically the first FLUX 3 open release, but it has nothing to do with making images.

FLUX 3 Image, the flagship still model

FLUX 3 Image handles text-to-image and editing through one endpoint. There's no mode switch: the prompt decides whether you're generating from scratch or changing an image you supplied. It accepts up to 10 reference images, and you can put bounding-box coordinates in the prompt (on a 0 to 1000 grid) to confine an edit to one region of the frame. Output runs from a 768-pixel square up to 4K, about 16 megapixels, across 15 aspect ratios.

Price climbs steeply with size:

A 4K image costs more than 12 times a 1K one. For social posts and most character work, 1K or 2K is the sensible setting. Keep 4K for print or for images you plan to crop hard.

On the Artificial Analysis text-to-image leaderboard, FLUX 3 Image sat at an Elo of 1117 on October 23, ranked somewhere between 8th and 15th. That's a solid mid-table placing. OpenAI's GPT Image 2.5 held the top spot at 1222.

FLUX 3 Image Klein, the one you can run yourself

Klein is a 5.6-billion-parameter version of FLUX 3 Image with the same request format: text-to-image and editing in one model, up to 10 references, output up to 4K. BFL released it on October 15 with weights on Hugging Face in two variants. Distilled runs in 16 steps and is the one to generate with. Base runs in 50 steps and exists for fine-tuning and LoRA training.

The license is the FLUX Kommunity License v1.0. Non-commercial use is free. Commercial use of the images is also free for "qualifying users", meaning businesses with under $5 million in gross annual revenue, affiliates included. That exception covers generating images; commercial fine-tuning and redistribution of the model need separate terms. Anyone bigger needs a paid self-hosting license or the API. That's tighter than the smallest FLUX.2 model, FLUX.2 [klein] 4B, which shipped under Apache 2.0.

Klein is also cheap through the API: $0.009 per image at 1K and $0.036 at 4K. On Artificial Analysis it scored 1043, the highest of any open-weights model on the text-to-image board.

Local hardware needs are modest by FLUX standards. BFL's launch post says an INT8 build runs on RTX cards with as little as 8GB of VRAM and makes a 1-megapixel image in about four seconds on an RTX 5090. ComfyUI added native support on launch day, with INT8 weights and a ready-made template, and measured under eight seconds on an RTX 5070 and under nine on an RTX 4070. Two caveats from BFL's own model card: rendered text can come out wrong, and the separate 9B prompt upsampler BFL pairs with Klein is a second model with its own memory cost. On an 8GB or 12GB card, plan to write fuller prompts yourself. Our guide to how much VRAM you need for local image and video models covers where Klein fits against SDXL, FLUX.2 and the local video models.

LoRA training worked from day one in AI Toolkit and NVIDIA NeMo, so expect community style and character LoRAs to show up quickly.

FLUX 3 Video and FLUX 3 Max

FLUX 3 Video has been in preview on BFL's API since August 4. It makes clips up to 20 seconds long at 24fps, with synced audio generated at the same time, including multilingual speech with lip sync, at no extra charge. It does text-to-video, image-to-video with one to 10 keyframes, and video-to-video (capped at 15 seconds). A draft mode renders at HD for cheap previews. QHD and UHD output arrived on September 10.

A 10-second FHD clip costs $2.90. A draft of the same shot costs $0.60, which makes drafting first the obvious habit. Video-to-video costs more at every tier.

FLUX 3 Max is a version of FLUX 3 that fal's research team post-trained with Black Forest Labs. fal serves it, BFL links to it from its own site, and it joined fal's catalog in mid-October. fal lists it at 720p for $0.17 per second, discounted to $0.0425 per second until November 7. fal says Max renders a 10-second clip in about 12 seconds and quotes a preliminary Artificial Analysis image-to-video score tied for first place. Those are fal's own figures. Krea carries Max too.

Where to run each FLUX 3 model

As of October 23, 2026:

• BFL API and Dev Playground: FLUX 3 Image, Klein and FLUX 3 Video.

• fal: all four, including Max, plus Klein's edit endpoint.

• Replicate: FLUX 3 Image, plus FLUX 3 Video in early access. No Klein.

• OpenRouter: FLUX 3 Image, Klein and FLUX 3 Video.

• Krea: FLUX 3 Image, FLUX 3 Video and Max.

• Runware: FLUX 3 Image, Klein and FLUX 3 Video.

• Hugging Face and ComfyUI: the Klein weights, for running locally.

What happened to FLUX.2 and the old Playground

Nothing has been switched off. FLUX.2 [pro], [max], [flex] and both [klein] sizes are still on BFL's API with no sunset dates, and FLUX.2 [dev] remains a local-only download. FLUX.1 Kontext and FLUX1.1 [pro] are listed as previous generation but still priced and available.

The consumer Playground at playground.bfl.ai is the one real casualty. It closed on August 31, 2026. Its replacement, the Dev Playground inside BFL's dashboard, covers text-to-image and multi-reference editing but dropped saved styles and presets, the community gallery and the Fill and Expand tools. Past generations didn't carry over, so anything left unexported is gone.

FLUX 3 FAQ

Is FLUX 3 open source?

Partly. Klein's weights are downloadable under the FLUX Kommunity License, which is free for non-commercial use and lets businesses under $5 million in revenue use its output commercially, but it isn't an open-source license. FLUX 3 Image and FLUX 3 Video are API-only. BFL says an open "FLUX 3 Dev" model is coming but hasn't given a date.

Can I use FLUX 3 Klein commercially for free?

For the images it generates, yes, if you or your business makes under $5 million in gross annual revenue, counting affiliates. Selling a fine-tuned version or redistributing the model needs separate terms. Above that revenue line, you need a self-hosting license from BFL or you can pay per image through the API.

How much VRAM does FLUX 3 Klein need?

BFL says the INT8 build runs on RTX cards with 8GB. In BF16 the generator alone takes 11.2GB, and BFL's docs describe it as running "down to an RTX 3090", a 24GB card. The prompt upsampler needs extra memory on top.

Is FLUX 3 Max an image model?

No. FLUX 3 Max makes video. It's a version of FLUX 3 that fal post-trained with Black Forest Labs and hosts, sold per second of output.

What replaced the BFL Playground?

The Dev Playground in BFL's dashboard took over when the old Playground closed on August 31, 2026. Existing accounts carried over, but saved styles, the community gallery and past generations did not.