Key details

  1. Released October 15, 2026 by Black Forest Labs.
  2. 5.6B image generator with base and distilled checkpoints.
  3. Up to ten reference images, bounding-box layout control and editing.
  4. INT8/NVFP4/FP8 builds and day-one Diffusers/ComfyUI support.
  5. Vendor reports 8GB GPU capability and workload-specific latency figures.
  6. FLUX Kommunity License: limited commercial output rights for qualifying businesses under $5M gross annualised revenue.

What builders should take away

  1. Read the actual FLUX Kommunity License before embedding the weights in a commercial service.
  2. Benchmark on your own GPU with the full text encoder and editing workload.
  3. Use distilled weights for latency-sensitive inference and base weights for supported customisation.
  4. Do not equate bounding-box guidance with pixel-perfect masking.
  5. Compare API cost, local hardware and legal rights before choosing deployment.

What changed

On October 15, 2026, Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weights image model with base and distilled checkpoints, BF16/FP8/INT8/NVFP4 builds, text-to-image, multi-reference editing and bounding-box layout controls. Hugging Face Diffusers and ComfyUI support arrived on launch day. The vendor says an INT8 build runs on an 8GB RTX GPU. The model is under the FLUX Kommunity License: qualifying smaller businesses can generate commercial outputs under stated conditions, while larger commercial deployments require separate licensing or API access.

Why it matters

Developers and small creative SaaS teams can evaluate a locally runnable image model with explicit placement/editing controls instead of relying only on hosted image APIs. Quantisation broadens hardware access and open weights enable experimentation, but the licence differs materially from permissive Apache-licensed models: commercial rights, fine-tuning, redistribution and larger-business use must be checked against the actual terms.

A 5.6B model with layout-aware editing

The model supports text-to-image, image editing and up to ten reference images. Bounding boxes in prompts can guide object placement or select editing regions; they are not hard pixel masks. The base checkpoint targets customisation and fine-tuning, while the distilled checkpoint defaults to 16 inference steps.

Hardware and benchmark context

Black Forest Labs reports that INT8 enables use on GPUs with 8GB VRAM and that its distilled model produces a 1MP image in roughly four seconds on an RTX 5090 and under nine on an RTX 4070. Hugging Face reports around 1.93 seconds for text-to-image on an H100 in its Diffusers test. These figures use different hardware and measurement setups; vendor quality rankings are not independently established.

Diffusers and local workflows

Hugging Face's launch guide documents ModularPipeline support, base and distilled checkpoints, quantised variants, LoRA training and image editing. GPU memory requirements depend on the full pipeline, text encoder, precision and offloading, not just the generator weight size.

The licence is a practical decision point

The FLUX Kommunity License permits non-commercial use and commercial output generation by qualifying users with less than US$5 million in gross annualised revenue, including affiliates, subject to conditions. This is not a blanket permissive software licence for all commercial fine-tuning, hosting or redistribution. Larger organisations need commercial licensing or API terms.

Hosted API versus local cost

Black Forest Labs also offers the distilled model through its API, advertising starting pricing around $0.009 per 1K image. Teams should verify current pricing, compute costs and licence rights rather than assuming self-hosting is always cheaper or unrestricted.

What to watch next

  • Independent quality and latency evaluations.
  • Licence clarifications for hosted SaaS, LoRA distribution and commercial derivatives.
  • Optimisations for lower-memory GPUs and broader runtime support.
  • Real-world use of layout controls in production creative tools.

Still unclear

  • Vendor quality comparisons and GPU latency claims are not independently replicated here.
  • The exact licence terms govern rights beyond the simplified revenue threshold.
  • Peak GPU memory varies substantially with pipeline and optimisation settings.

Sources

Direct reading behind this dossier.

3 sources
FLUX 3 Image Klein
Black Forest Labs primary model launch

Model features, hardware claims, API pricing and commercial licence summary.

Discussion

Discussion is reader-contributed. Comments are not part of the BTN dossier or its editorial evidence.

0 visible comments

Join the discussion

Keep comments useful and relevant. Reader contributions may be moderated and are not BTN editorial evidence.

Sign in to comment