FLUX Image Models Explained: Open Weights, Licences, Pricing and API

FLUX is a family of image (and now video) generation models from Black Forest Labs, released partly as open weights and partly as API-only models, and each tier carries a different licence, price and deployment route. Black Forest Labs (BFL) announced itself and the FLUX.1 suite on 1 August 2024, followed by FLUX.2 on 25 November 2025, FLUX 3 in early access on 23 July 2026 and the open-weight FLUX 3 Image Klein on 15 October 2026. This page sets out every current model tier, which weights are open and under which licence, the API and its per-image prices, self-hosted commercial licensing, provenance and safety controls, data handling and the dated release history. It is written for developers, product and creative teams, platform engineers and the legal, privacy and security teams who approve image generation tools.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Three licence routes. Some FLUX weights are Apache 2.0 (FLUX.1 [schnell], FLUX.2 [klein] 4B), some are non-commercial ([dev] models, FLUX.2 [klein] 9B), and FLUX 3 Image Klein uses the FLUX Kommunity License.
• USD 5M threshold for FLUX 3 Image Klein. The Kommunity License allows free commercial use for individuals and companies with annual revenue under USD 5M.
• Pro, Max and Flex tiers are API-only. FLUX.2 [pro] and [flex] are managed models served through the BFL API, Playground and launch partners.
• Priced per image or per megapixel. FLUX 3 Image Klein starts at USD 0.009 for a 1K image; rejected, failed or moderated outputs are not charged.
• Provenance on the API. BFL states the FLUX.2 Pro API applies cryptographically signed C2PA metadata, and its open-weight reference code includes a pixel-layer watermarking example.

What is FLUX?

FLUX is a family of rectified flow transformer models for generating and editing images from text and reference images. FLUX.1 [dev] and [schnell] are described on their model cards as 12 billion parameter rectified flow transformers. FLUX.2 uses a latent flow matching design that pairs a Mistral-3 24B vision-language model with a rectified flow transformer. BFL's documentation now describes FLUX as models that generate and edit images and video, render synchronised audio and drive robotics, with FLUX 3 presented as "one model, one API". BFL describes its founding team as researchers whose prior work includes VQGAN, Latent Diffusion and Stable Diffusion-related models. Its launch announcement reported a USD 31 million seed round led by Andreessen Horowitz.

Key facts

Models and versions

As of 2026-11-15, from bfl.ai/models, BFL announcements and Hugging Face model cards. FLUX 3 Dev is announced as a planned open-weight model with no date.

Naming: the bracketed suffix marks the tier. [schnell] (German for "fast") is distilled to generate in about 1 to 4 steps; [dev] is guidance-distilled for efficiency; [klein] ("small") is size-distilled for low latency; [pro], [flex] and [max] are managed API tiers.

Pricing and plans

As of 2026-11-15, from bfl.ai/pricing. Tier sizes are approximate: 768 is about 0.6 MP, 1K about 1 MP, 2K about 4 MP and 4K about 16 MP. Reference images on FLUX.2 are billed per megapixel. Self-hosted licences are priced by tier or quote. Running open weights yourself has no per-image fee, only hardware and operating cost.

Are FLUX models open source, and how do the licences work?

"Open weights" is not the same as open source for every FLUX model. The weights for several models are downloadable, but only some are under a permissive licence.

As of 2026-11-15. BFL's commercial tiers, from bfl.ai/licensing:

• Builder: FLUX 3 Image Klein and FLUX.2 [klein]; 10K images a month, one domain, ten licensed users; fine-tuning and LoRA rights; not for client use or downstream applications.
• Platform: FLUX 3 Image Klein, FLUX.2 [klein] Base 9B and FLUX.2 [dev]; 100K images a month for SaaS and product teams.
• Professional: FLUX 3 Image Klein and FLUX.2 [dev]; 100K images a month across up to three client domains, with per-client fees after three.
• Enterprise: all models and new releases on custom terms, including combined API and weights pricing.
• Synthetic Data: training rights on outputs, scoped to the contracted use case.

The standard licence allows self-hosting, fine-tuning, LoRAs and integration into products, but not an external-facing API endpoint serving the model. Builder and Platform terms are standard; Professional and Enterprise can be negotiated. See the Open-Weight vs Closed Models Explained for the general distinction.

Features

• Multi-reference generation: FLUX.2 and FLUX 3 Image combine up to ten reference images, each referred to by its position in the prompt.
• Resolution: FLUX.2 generates and edits up to 4 MP; FLUX 3 Image Klein offers native 4K generation.
• Editing: instruction-based editing, with bounding boxes on a 0 to 1000 grid to place elements or confine edits (FLUX 3).
• Grounding: FLUX 3 Image can search the web for real-world details; the grounding parameter defaults to on and can be set to false for faster, prompt-only results.
• API extras: pixel restoration of unedited regions and prompt upsampling for FLUX 3 Image Klein.
• Tools: outpainting, erase, virtual try-on and deblur endpoints, plus video upscaling and editing for FLUX 3 Video.

Availability

Hardware, as stated by BFL: FLUX 3 Image Klein generator weights are 11.2 GB in BF16 and 5.6 GB in FP8, and the INT8 build runs on NVIDIA GeForce RTX GPUs with as little as 8 GB VRAM. FLUX.2 [klein] 4B needs about 13 GB VRAM, with the RTX 3090 and 4070 given as examples.

A practical self-hosting sequence:

1. Pick the model by licence first: Apache 2.0 for unrestricted commercial use, Kommunity if under USD 5M revenue, otherwise a Commercial Weights tier.
2. Accept the gated licence on Hugging Face with an organisation account and record who accepted it.
3. Download from the black-forest-labs organisation only, verify hashes and pin library versions.
4. Keep the reference code's NSFW and IP filters and watermarking enabled; the non-commercial licence requires filters or manual review.
5. Run inference in an isolated environment with authenticated access; the standard licence does not permit exposing the model as an external API.
6. Add C2PA signing or labelling in your pipeline and log usage under your retention policy.

See the Open-Weight Model Security Guide and the Hugging Face Security Guide for model supply chain and token controls.

Which FLUX route fits which need?

The table maps common requirements to the routes BFL documents. It sets out factual criteria, not a recommendation.

Small teams usually face the simplest choice: an Apache 2.0 or Kommunity model, or pay-per-image API calls. Larger organisations and agencies need to match image volumes, domains and client counts to a Commercial Weights tier, and should confirm with BFL in writing where a use case sits between tiers.

Provenance and labelling

The FLUX.2 [dev] model card states that the FLUX.2 Pro API applies cryptographically signed C2PA metadata to outputs, and that the open-weight inference code implements an example of pixel-layer watermarking. The FLUX.2 [klein] 4B card lists pixel-layer watermarking and C2PA metadata support. For the API, BFL describes multiple filters on prompts, uploaded images and outputs, including third-party filters from Hive and Microsoft, with some categories adjustable by developers. Under the non-commercial licence, deployers must use filters or manual review, and BFL may spot-check. Self-hosted outputs carry only what your pipeline keeps, and C2PA metadata is lost through many re-encoding and upload paths. See the AI Content Provenance and C2PA Guide.

Data and privacy

• Training use: BFL's privacy policy says prompts, uploaded images and outputs may be used to train and improve models, after safeguards such as data minimisation and de-identification.
• Opt-out: email [email protected] with the subject "Training Opt Out" and account details.
• Retention: no fixed period; data is kept as long as reasonably necessary to provide the service and for safety, dispute and legal purposes.
• Controller: BFL GmbH for Switzerland, the UK and the EEA; Black Forest Labs Inc. elsewhere. Transfers rely on standard contractual clauses.
• Enterprise and third parties: enterprise content under contract is governed by that contract; via a third-party host, that host is the data controller.
• API keys: BFL advises keeping the key in an environment variable and never in client-side code; store it in a secrets manager and rotate it. See the Self-Hosted LLM Security Guide for comparable self-hosting controls.

Release history

As of 2026-11-15, from bfl.ai announcements and blog. Release dates for FLUX 1.1 [pro], FLUX.1 Kontext and FLUX.2 [klein] are not stated on the BFL pages checked.

Limits

• Not factual: model cards state the models cannot provide factual information and may not match prompts.
• Bias: the cards note the models may reflect and amplify societal biases.
• Prohibited uses: out-of-scope uses include non-consensual nudity, harm to minors, disinformation, harassment and fully automated decisions affecting legal rights.
• Early access: FLUX 3 capabilities roll out after early access phases, and BFL describes its FLUX 3 evaluations as preliminary.
• Licence scope: a non-commercial or Builder licence does not cover client work or serving the model to others.

Practitioner checklist

• Record each FLUX model in use with its exact licence (Apache 2.0, Non-Commercial, Kommunity or a commercial tier).
• Confirm revenue eligibility before relying on the Kommunity License, and buy a commercial tier where needed.
• Download weights only from the official black-forest-labs Hugging Face organisation and verify hashes.
• Keep safety filters and watermarking in the reference code switched on for self-hosted deployments.
• Send the training opt-out email where hosted inputs must not train models, or use an enterprise contract.
• Store BFL API keys in a secrets manager, never in client-side code, and rotate them.
• Preserve C2PA metadata from API outputs and label self-hosted outputs where disclosure rules apply.
• Track BFL announcements for new tiers and pricing changes, and re-check per-image costs before budgeting.

Standards and references

• Black Forest Labs: Models
• Black Forest Labs: Pricing
• Black Forest Labs: Licensing
• Black Forest Labs: FLUX 3 Image Klein
• Black Forest Labs: FLUX 3, Real World Models
• Black Forest Labs: FLUX.2
• Black Forest Labs: Announcing Black Forest Labs
• Black Forest Labs: FLUX.2 [dev] model card
• Black Forest Labs: FLUX.1 [schnell] model card
• Black Forest Labs: API quick start
• Black Forest Labs: Privacy Policy

Related NHI Mgmt Group resources: Open-Weight Model Security Guide · AI Content Provenance and C2PA Guide · AI Image and Video Generation at Work · Hugging Face Security Guide