Compare FLUX.2 pro, max, flex, dev and klein. Learn reference editing costs, hardware limits, licensing, and a practical image-generation workflow.
FLUX is Black Forest Labs’ family of image-generation and editing models. It offers hosted APIs and downloadable weights, giving teams a choice between managed access and running a model on their own infrastructure.
The practical questions are which variant fits your brief, what reference images cost, and whether its license permits your deployment. This guide focuses on FLUX.2 images rather than the separate video lineup. Visit Black Forest Labs or try the BFL Playground.
FLUX.2 launched on November 25, 2025. It combines text-to-image generation with image editing and multiple references. BFL positions the family for photographic imagery, product visuals, layout-sensitive designs, and workflows requiring subject consistency.
References help establish a product, character, or style, but do not guarantee exact preservation. Inspect labels, proportions, and small details after every edit.
BFL’s model-selection guide distinguishes these roles:
| Variant | Main reason to consider it | Access |
|---|---|---|
| FLUX.2 [pro] | Balance of quality, speed, and cost | Hosted API |
| FLUX.2 [max] | Premium output and optional grounding search | Hosted API |
| FLUX.2 [flex] | Adjustable steps and guidance | Hosted API |
| FLUX.2 [dev] | Local experiments and customization | Open weights |
| FLUX.2 [klein] | Compact, low-latency generation and editing | API or open weights |
Calling pro the highest-quality variant overlooks max. Flex is useful when you want to adjust sampling rather than accept a fixed setting.
Reference limits also differ: BFL lists up to four for klein, up to six for dev, and eight to ten for pro, max, and flex depending on access route. Check the exact endpoint before building a workflow around a particular limit.
Grounding search can help max incorporate current information, but a generated image should not be treated as evidence that a depicted fact is correct.
The klein announcement describes sub-second inference and consumer-GPU support. Those are separate claims: it does not mean every consumer GPU generates an image in half a second.
BFL states that its speed comparison uses GB200 hardware in bf16. The 4B model can fit in roughly 13GB VRAM, while quantized versions reduce memory needs. Your latency depends on hardware, precision, image size, and software.
Distilled klein models prioritize short inference runs. Base variants retain more sampling flexibility for fine-tuning and experimentation, so do not assume they share the same speed.
The official pricing page, checked October 8, 2026, lists:
| Variant | First output MP | Additional output MP | Reference input per MP |
|---|---|---|---|
| klein 4B | $0.014 | $0.001 | $0.001 |
| klein 9B | $0.015 | $0.002 | $0.002 |
| pro | $0.030 | $0.015 | $0.015 |
| max | $0.070 | $0.030 | $0.030 |
| flex | $0.050 | $0.050 | $0.050 |
BFL defines one MP here as 1024 × 1024 pixels and rounds input and output sizes up separately. Multiple references are each billed as one MP under the listed rules.
For example, pro generating a one-MP output with two one-MP references costs $0.03 + $0.015 + $0.015 = $0.06, before other applicable charges. A headline starting price is not the complete editing cost.
Klein 4B and 4B Base use Apache 2.0. Klein 9B variants use the FLUX Non-Commercial License. Do not describe every open-weight FLUX model as unrestricted open source.
The FLUX.2 dev model card states that generated outputs can be used commercially as described by its license. That is different from permission to deploy the model commercially.
The non-commercial license and BFL’s commercial licensing terms govern model operation. A business deploying dev inside a paid service must review that deployment’s requirements separately from using a generated asset. Hosted API access has its own terms.
Local execution also has hardware and operating costs. Downloadable weights do not make production inference cost-free.
The Wiro ML Team’s April 2026 test used klein Base 4B for five 1024 × 1024 outputs at 30 steps and guidance 4.0.
The team found useful product-photography cues, while emphasizing inspection of close details and generated lettering. This is a concrete example of an operating configuration, not proof of universal quality.
Wiro also provides model access, so this is a provider-authored hands-on assessment rather than an independent benchmark. Five samples cannot establish performance across the whole FLUX family. This guide does not claim firsthand testing.
A practical prompt is:
> A ceramic teapot steaming on a rustic wooden table, soft morning window light, editorial food photography, shallow depth of field, muted earth tones, uncluttered background.
Put important subjects first and describe observable details. There is no universal ideal word count; add instructions when they resolve ambiguity.
FLUX is worth evaluating when deployment flexibility and reference-led editing matter. Choose the model using a real brief, budget for revisions and references, and keep deployment licensing separate from output rights.
FLUX is a family of image-generation and editing models available through hosted APIs and downloadable weights. This guide focuses on the FLUX.2 image family.
Consider pro for balanced production work, max for premium output and grounding search, flex for sampling controls, klein for compact low-latency workflows, and dev for local customization.
No. BFL's advertised speed measurements depend on specific hardware and settings. Consumer-GPU compatibility does not guarantee the same latency.
No. Klein 4B and 4B Base use Apache 2.0, while klein 9B variants and FLUX.2 dev use the FLUX Non-Commercial License.
Its model card permits commercial use of generated outputs as described by the license. Commercial deployment of the model is a separate question and must be reviewed under the applicable terms.
Yes. Reference inputs are billed in addition to output resolution. For pro, a one-MP output with two one-MP references totals $0.06 at the rates checked October 8, 2026.
Yes. FLUX.2 supports generation and image editing with references. Reference limits differ by model and access route, so check the endpoint documentation.
Some weights can be downloaded without a model purchase, subject to their license. Local generation still requires suitable hardware and has operating costs.
