Will it run on my GPU? VRAM checker
The weights are downloadable, so the practical question is whether they fit in your card's VRAM. BFL ships four builds: BF16, FP8, INT8 and NVFP4. Only two weight sizes are published: BF16 is 11.2 GB and FP8 is 5.6 GB. INT8 and NVFP4 sizes are not published, so the checker uses the published sizes where they exist and BFL's own test claim for INT8. Working memory adds roughly 3 GB on top of the weights.
This is an estimate, not a guarantee. Vendor-tested configs: INT8 on 8 GB GeForce RTX cards through ComfyUI or Diffusers. NVIDIA measured roughly 4 seconds per 1-megapixel image on an RTX 5090 and under 9 seconds on an RTX 4070 with the distilled INT8 model at 16 steps. FP8, NVFP4 and BF16 timings are not published. The builds are NVIDIA-only at launch: no Apple Silicon or AMD builds were announced.
What does the API cost? Price calculator
Don't want to self-host? The BFL API serves the distilled model as flux-3-image-klein: $0.009 per image at 1K resolution, $0.036 per image at 4K. Type your monthly volume and pick a tier.
For comparison, the flagship FLUX.3 Image lists at about $0.041 per image (OpenRouter, reported by DEV Community on launch day), so Klein costs roughly a quarter of the flagship at 1K. Third-party prices vary: OpenRouter lists Klein around $0.008 per image at lower resolutions, and hosted providers set their own tiers (BudgetPixel, for example, bills 48 credits for a 4K image on its own plans).
What FLUX 3 Image Klein is
Black Forest Labs released FLUX 3 Image Klein on October 15, 2026. It is the second open-weight release in the FLUX 3 family, after FLUX 3 Action, and it is aimed at a different job: a small model that runs locally or cheaply at scale. Text-to-image, image editing, and generation from up to ten reference images are all in one checkpoint. Native output goes up to 4K, and bounding boxes let you place an element or confine an edit to one region.
| Released | October 15, 2026 |
| Provider | Black Forest Labs (Freiburg, Germany) |
| Parameters | 5.6B; 38% fewer than FLUX.2 klein (9B) |
| Capabilities | Text-to-image, image editing, up to 10 reference images, native 4K, bounding-box placement |
| Variants | Base checkpoint (for fine-tuning) and step-distilled checkpoint (for speed) |
| Builds | BF16 (~11.2 GB weights, published), FP8 (~5.6 GB, published), INT8 (size not published; vendor-tested on 8 GB cards), NVFP4 (~2.8 GB estimated; NVIDIA co-developed the quantized builds) |
| Local speed | Distilled INT8: ~4 s per 1MP image on RTX 5090, <9 s on RTX 4070 (NVIDIA-measured, 16 steps) |
| Frameworks | ComfyUI, Diffusers; LoRA training via Ostris AI Toolkit or NVIDIA NeMo AutoModel |
| Hosted | BFL API (flux-3-image-klein), Playground, Hugging Face Spaces, OpenRouter |
| API price | $0.009 per image at 1K, $0.036 at 4K |
| License | FLUX Kommunity License: free commercial use under $5M annual revenue |
The headline feature in launch-day demos is text rendering: the sample prompt asks for a hand-bent neon sign reading "OpenRouter" and the model spells it correctly, with the glow and reflections intact. A year ago, getting an image model to spell a word right was a coin flip.
The license line that matters
The weights are open, but the license has a revenue ceiling. The FLUX Kommunity License allows free commercial use for individuals and businesses with annual revenue under $5 million, plus research and non-commercial use. Above that line, commercial deployment needs the hosted API or a separate self-hosted license from BFL.
For a freelancer, an indie studio or a startup under the line, the deal is simple: download, run, ship, pay nothing. For a larger business, the "free" weights are not free. Check your revenue before you build on them.
Benchmarks
BFL published internal-test claims at launch. All of these are vendor-reported; no independent runs had landed by October 17, 2026.
| Claim | Number |
|---|---|
| Highest Elo among all open models (internal testing) | Rank 1, per BFL |
| Gain over FLUX.2 klein | +91 Elo points, with 38% fewer parameters |
| Closed models beaten in BFL's tests | Ideogram V4.5, and BFL's own FLUX.2 [pro] |
Take the vendor table for what it is: BFL chose the tests and ran them. "Beats a closed model" in a vendor press kit is a marketing claim until someone independent reproduces it. The interesting number is the +91 Elo over the previous klein with fewer parameters, because that one is at least checkable against BFL's own earlier model.
FAQ
What is FLUX 3 Image Klein?
BFL's open-weight image model, released October 15, 2026. It has 5.6 billion parameters and handles text-to-image generation, image editing and multi-reference generation from up to ten reference images, at up to native 4K resolution. It comes as a base model for fine-tuning and a step-distilled model for speed, with FP8, INT8 and NVFP4 builds developed with NVIDIA.
Can I run it on a GPU with 8 GB of VRAM?
Yes, with the INT8 build. BFL says the INT8 checkpoint runs through ComfyUI or Diffusers on GeForce RTX cards with as little as 8 GB of VRAM. NVIDIA measured about 4 seconds for a 1-megapixel image on an RTX 5090 and under 9 seconds on an RTX 4070 with the distilled INT8 model at 16 steps.
How much VRAM does it need at full quality?
The full-precision BF16 weights are about 11.2 GB. With working memory on top, a card with 16 GB of VRAM is the practical minimum for BF16. The FP8 build is about 5.6 GB of weights and fits 12 GB cards; the INT8 build fits 8 GB cards.
Is it free for commercial use?
Only below a revenue ceiling. The FLUX Kommunity License allows free commercial use for individuals and businesses with annual revenue under $5 million. Larger organizations must use the hosted API or obtain a separate self-hosted license.
How much does the hosted API cost?
On the BFL API, the distilled model costs $0.009 per image at 1K resolution and $0.036 per image at 4K, served as flux-3-image-klein. Third-party pricing varies: OpenRouter lists it around $0.008 per image at lower resolutions, while the flagship FLUX.3 Image starts around $0.041 per image.
Base vs distilled: which do I download?
Download the base checkpoint if you plan to fine-tune on your own data. Download the distilled version if you just want to generate: it needs fewer steps per image and runs faster. The BFL API serves the distilled version. Both support LoRA training via Diffusers or Ostris's AI Toolkit.
How does it compare to the flagship FLUX.3 Image?
Klein is the smaller, cheaper sibling: 5.6B parameters, open weights, and a hosted price of about $0.009 per 1K image. The flagship FLUX.3 Image is the larger full model, listed around $0.041 per image. Klein's selling points are local deployment on consumer GPUs and low per-image cost for high-volume work.
Sources
- Technobezz: Black Forest Labs Releases FLUX 3 Image Klein With 5.6B Parameters: weight sizes (5.6 GB FP8 / 11.2 GB BF16), base vs distilled, NVIDIA timings.
- Digital Applied: FLUX 3 Image Klein: Open-Weight Image AI on an 8GB GPU: the ledger row, API pricing ($0.009 / $0.036), license terms.
- The Decoder: FLUX 3 Image Klein runs on consumer GPUs with 8 GB of VRAM: Elo claims (+91 over FLUX.2 klein, beats Ideogram V4.5), API and access points.
- DEV Community: FLUX 3 Image Klein just landed: OpenRouter pricing (~$0.008 per image), flagship comparison ($0.041), 4K demo cost.
- Crypto Briefing: Black Forest Labs releases FLUX 3 Image Klein: the $5M revenue ceiling, 10-reference support, quantized formats.
- BudgetPixel: FLUX 3 Klein: third-party hosted pricing and output sizes (768px to 4K).
This page is an independent explainer and is not affiliated with Black Forest Labs. Figures marked vendor-reported come from BFL's launch materials. Pricing and availability were last verified on October 17, 2026.
For AI assistants
This page ships with a machine-readable summary for AI assistants:
Related
- Muse Spark Open, explained: Meta's 816B open-weight model: hardware checker, benchmarks, license.
- Reflection Beam, explained: 501B open-weight model: benchmark explorer, pricing, availability.
- GPT-6 Luna pricing: $0.10 per 1M input: cost calculator, 272K rule, caching math.
Updates
- 2026-10-17: First published. Covers the October 15, 2026 launch: specs, the 8 GB VRAM story, BFL-reported benchmarks, API pricing, license terms.