The FLUX.2 family
FLUX.2 is the only image family in this reference whose weights are downloadable and deployable on local infrastructure. That statement does not hold for every version though: of the six active versions, three ship published weights and three are API-only, and even among those three, the licences are not identical.
- klein 4B under Apache 2.0
- rate from $0.014
- per-megapixel rate structure
Last checked: This is a reference page. It is re-checked against the vendor sources and updated when a new version ships. Change log
A licence that changes with model size
This parameter is visible in Black Forest Labs official table and gets lost in most summaries: klein ships in two sizes under two distinct licences. klein 4B is published under Apache 2.0, meaning unrestricted commercial use is permitted. klein 9B is published under the FLUX Non-Commercial License, meaning the same weights are downloadable but cannot be used for client work. The dev model also carries published weights, and its licence is non-commercial as well.
Now place this benchmark data beside the licences. On the Arena text-to-image leaderboard, klein 9B scores 1069.6 and klein 4B scores 1030.1. The version with the more open licence also performs worse. If your workload is commercial, the choice is already constrained, and the 1030 score is what has to be worked with. This is the only point in this reference where a legal document reaches directly into output quality.
The rate structure here is per megapixel, not per image
Google and OpenAI charge a fixed rate per size tier. Black Forest Labs does not use this structure: cost rises with output resolution. So the $0.07 rate for max is the price of a one-megapixel image, and the same image at 4 megapixels costs four times as much.
For klein, the formula is published officially: the first megapixel carries a flat rate, and each additional megapixel adds to it. klein 4B computes as $0.014 plus $0.001 per additional megapixel, and the vendor own worked example places a 2MP image at $0.015. klein 9B follows the same structure at $0.015 plus $0.002. The figure logged in our category table is that published floor, so any output larger than one megapixel will bill above the table figure.
A contradiction inside their own official documentation
For flex, the family overview page logs a rate of $0.06 per megapixel in its "which model should I choose" table, while the pricing page in the same documentation set logs "from $0.05." These two figures are not identical, and this analysis has not selected either as final. Both documents were checked today, and both addresses are logged at the top of this page. For budgeting flex, the official rate calculator should be checked directly, and neither figure here should be relied upon.
Which version belongs in deployment
If the workload is commercial and must run locally, exactly one option is available: klein 4B. This version runs on an ordinary consumer graphics card with roughly 13GB of memory, uses four inference steps, and returns output in under a second. It is built for high-volume, real-time workloads, and its leaderboard score reflects that same constraint: it is low.
If deployment stays on the API, pro is the architectural balance point, running at half the rate of max. One parameter rarely documented: pro is the only version where editing an image costs more than generating one, $0.045 against $0.03. On every other version these two figures are identical.
max should be deployed when web search is required. It is the only version in this family that can search the web during generation, a parameter that matters directly for any image that must reference something real and current. flex, per official documentation, is specialized for typography, with adjustable step and guidance parameters.
If the objective is retraining the model
The Base variants should be selected, not the standard variants. Black Forest Labs publishes two separate weight families for klein: the distilled version, which is fast, and the Base version, which is undistilled and carries higher output diversity. Official documentation states Base is intended for fine-tuning, LoRA training, and research, and is not offered on the public API. The licences carry over unchanged: 4B Apache, 9B non-commercial.
Two parameters that consume disproportionate time
First, the numbering. FLUX 3 is newer than FLUX.2 but is not its successor: FLUX 3 is the video production line and is priced per second, while FLUX.2 is the image line. If the workload is images, the larger version number is not the parameter needed.
Second, endpoints. For pro and klein 9B, two separate endpoints exist: one with a preview suffix, where the latest improvements are deployed, and one without a suffix, which is a fixed snapshot of the model that does not change. If output must remain identical across runs, the unsuffixed endpoint should be called. This is a distinction that goes unnoticed until the day output shifts for no apparent reason.
Access status from Iran
Unknown, and this time the reason differs from Google and OpenAI. Those two vendors publish supported-countries lists, and this analysis is written from those same lists showing Iran absent. Black Forest Labs has no such document: its complete documentation index was checked, and it contains neither a terms page nor a country list, and the terms URL on the main site returns a 404 status. So the Iran field for this family is logged as empty, and that empty state means no document existed to check, not that no check was performed.
Specifications
Every number here comes from the vendor's own page, with that page linked beside it.
Release timeline
Each version with its own release date, and what changed against the one before it.
- FLUX.2 [klein] 4B current
- FLUX.2 [max] current
Documented strengths
- The only image family in this reference with downloadable weights, and klein 4B is published under Apache 2.0
- klein 4B runs on an ordinary consumer GPU with roughly 13GB of memory
- The lowest rate floor among all candidates in the image category: from $0.014
- max is the only model in this family that searches the web during generation
- Undistilled Base variants are published for fine-tuning and LoRA training
- A fixed endpoint is available for cases where output must be reproducible across runs
Known weaknesses and limits
- No version in this family ranks near the top of either Arena leaderboard; the best-ranked sits at 24.
- Open weights and top-tier quality do not deploy together: the 9B version with the higher benchmark score carries the non-commercial licence.
- Rate is per megapixel, so the table figure is a floor, not the final rate for a large image.
- Official documentation logs two different rates for flex: $0.06 in one document and from $0.05 in another.
- Editing with pro costs more than generating with pro, a parameter not visible in the primary rate column.
- No document on permitted countries is published, so the Iran field for this family is empty.
- No separate resolution ceiling is stated for klein; the 4-megapixel figure describes the whole family.
Technical verdict
If the workload is commercial and must be deployed locally, klein 4B is the only real option in this table, and its quality gap from the top of the table has to be accepted. If deployment stays on the API, pro should be selected, and max only when web search is required. And for flex, until these two figures agree inside their own official documentation, the rate should be extracted directly from their official calculator.
Frequently raised questions, with documented answers
Is FLUX.2 free
klein 4B weights are published under Apache 2.0, and running them locally costs the vendor nothing. On the API, the rate starts at $0.014. The dev model also has published weights, but its licence is not commercial.
Which FLUX.2 version should be downloaded
For commercial workloads, only klein 4B, since it is the only version carrying an Apache 2.0 licence. klein 9B and dev have accessible weights but non-commercial licences.
What hardware does FLUX.2 require
Official documentation logs roughly 13GB of graphics memory for klein 4B, equivalent to an ordinary consumer card. This version uses four inference steps and returns output in under a second.
What is the difference between FLUX.2 and FLUX 3
FLUX 3 is not the successor to FLUX.2. FLUX 3 generates video and is priced per second; FLUX.2 is the image production line.
What does one FLUX.2 image cost
It depends on output size, since rate is structured per megapixel. The floors: klein 4B from $0.014, pro from $0.03 to generate and $0.045 to edit, and max from $0.07.