Black Forest Labs' FLUX 3 Klein Hits Top-2 Open Image AI at $0.009
Black Forest Labs' 5.6B open-weights model lands within one Elo point of Qwen-Image-2.1 on text-to-image while costing a fraction of larger rivals.
- FLUX 3 Image Klein ranks #2 open-weights on Artificial Analysis T2I at Elo 1043, one point behind Qwen-Image-2.1.
- 5.6B parameters, runs on 8GB VRAM consumer GPUs with INT8 quantization.
- Priced at $9 per 1,000 images on BFL API, 40% cheaper than FLUX.2 [klein] 9B.
- Free commercial use for individuals and companies under $5M annual revenue under FLUX Kommunity License.
- Ships base checkpoint, step-distilled variant, plus FP8, INT8, NVFP4 builds optimized with NVIDIA.
- Biggest gains: Composition & Framing, Layout, Knowledge, and UI/UX Design use cases.
Black Forest Labs’ 5.6-billion-parameter FLUX 3 Image Klein has reached second place among open-weight text-to-image models on Artificial Analysis. Its Elo score of 1,043 trails Qwen-Image-2.1 by one point, within the evaluation’s margin of error. The model also costs $0.009 per 1K-resolution image through Black Forest Labs’ API, about 19% of the $0.048 price for the full FLUX 3 Image. Developers get top-three benchmark performance, consumer-GPU deployment, and a low API price under a revenue-limited license.
One GPU, several deployment paths
The release details describe a model for text-to-image generation, image editing, and generation from as many as 10 reference images. Klein also supports 4K output, localized editing, and bounding boxes that place generated elements or restrict edits to a selected area.
| Format | Intended use |
|---|---|
| Base weights | Fine-tuning and workflows that need full control over inference |
| Step-distilled model | Faster generation with fewer inference steps |
| FP8, INT8, and NVFP4 builds | Lower memory use and faster inference on supported hardware |
Quantized formats store weights at lower numerical precision, reducing VRAM requirements and often improving throughput. They can also change output quality, so production tests should use the same precision and inference settings planned for deployment.
Using the distilled INT8 model through ComfyUI, Black Forest Labs reports about four seconds for a one-megapixel image on an NVIDIA GeForce RTX 5090 and less than nine seconds on an RTX 4070. The company says the INT8 build can run with 8GB of VRAM. Those figures are vendor benchmarks; latency varies with image dimensions, step count, batch size, software stack, and GPU configuration.
One Elo point from first
In the reported leaderboard snapshot, Klein sits one point behind the text-to-image leader and third on image editing. An Elo score summarizes relative model preference in evaluation matchups and has no direct percentage interpretation.
| Benchmark | Rank | Model | Elo |
|---|---|---|---|
| AA-Image-T2I v2.1 | 1 | Qwen-Image-2.1 | 1,044 |
| AA-Image-T2I v2.1 | 2 | FLUX 3 Image Klein | 1,043 |
| AA-Image-T2I v2.1 | 3 | Fibo 2.0 | 1,040 |
| AA-Image-Editing v2.1 | 1 | Fibo 2.0 | 1,086 |
| AA-Image-Editing v2.1 | 2 | Qwen-Image-2.1 | 1,080 |
| AA-Image-Editing v2.1 | 3 | FLUX 3 Image Klein | 1,070 |
Compared with FLUX.2 [klein] 9B, the new model gains about 114 Elo points on text-to-image generation and 55 points on editing while using 38% fewer parameters. At the same numerical precision, that reduction also cuts weight storage by roughly 38%, although activations and other runtime data still contribute to total VRAM use.
Diagrams rise, faces lag
Artificial Analysis reports gains over the FLUX.2 lineup across all 19 text-to-image categories, with several areas showing especially large changes:
- Composition and framing: Cropping, reframing, and perspective changes show the largest gains among editing actions for both FLUX 3 models.
- Layout and knowledge: Klein moves from the bottom third to the top quarter, improving its position for diagrams, flowcharts, and prompts that depend on factual relationships.
- Animation and gaming: This is the highest-ranked use case for Klein and the full FLUX 3 Image model.
- UI and UX design: The full FLUX 3 Image model moves from the bottom half into the top five. Klein’s results should be evaluated separately for interface work.
Identity-preserving editing remains a relative weakness, particularly when a workflow must keep faces or characters consistent across revisions. Klein also scores lower on retail and e-commerce prompts. Some replies to Black Forest Labs’ launch post favor Qwen-Image-2.1 for speed and output quality in direct comparisons. Hardware, quantization, sampler settings, and prompt selection can account for part of that variation.
Low API price, limited license
| Model | Price per 1K image | Price per 1,000 images |
|---|---|---|
| FLUX 3 Image Klein | $0.009 | $9 |
| FLUX 3 Image | $0.048 | $48 |
Klein’s API rate is 40% below FLUX.2 [klein] 9B and about 81% below the full FLUX 3 Image model. Self-hosting shifts the expense to GPU time, engineering, storage, and operations, making cost per accepted image more useful than the listed inference price alone.
The FLUX Kommunity License permits free commercial use by individuals and businesses with annual revenue below $5 million. Larger organizations can negotiate self-hosting rights or use the hosted API. The revenue threshold makes the license narrower than permissive terms such as Apache 2.0. Deployments involving redistribution, fine-tuned weight distribution, or hosted services require review of the current license text.
Choose by failure mode
| Requirement | Assessment | Reason |
|---|---|---|
| Single-GPU self-hosting | Strong fit | The INT8 build reportedly runs with 8GB of VRAM. |
| Low-cost API generation | Strong fit | List pricing is $9 per 1,000 images at 1K resolution. |
| Multi-reference workflows | Strong fit | The model accepts as many as 10 reference images. |
| Diagrams and game assets | Promising | Layout, knowledge, animation, and gaming rank among its stronger categories. |
| Consistent faces or characters | Requires comparison testing | Identity-preserving edits remain a relative weakness. |
| Retail product imagery | Requires comparison testing | Retail and e-commerce prompts score below the model’s stronger categories. |
| Self-hosting above $5 million in revenue | Commercial agreement required | The free commercial grant ends at the stated revenue threshold. |
Teams comparing Klein with Qwen-Image-2.1 and Fibo 2.0 can run a fixed suite of production prompts using matched dimensions, reference images, hardware, and quality settings. Useful measurements include prompt adherence, text rendering, identity drift, latency, peak VRAM, retry rate, and cost per accepted image. The close Elo rankings leave workload-specific failures as the deciding evidence.
Comments
