Flux LoRA in 2026: Which Base Model You Can Legally Train On

Quick answer

• A Flux LoRA is a “Derivative” under Black Forest Labs’ licence, and the licence says so in as many words: a Derivative includes any “customized or fine-tuned version” of the model. Train on a [dev] model and the LoRA inherits the non-commercial restriction.
• The images are a separate matter. Outputs are explicitly not Derivatives, and you “may use Output for any purpose (including for commercial purposes)”. That is not a loophole for paid work, though — the two clauses have to be read together, and they are below.
• If you need a commercially usable LoRA, the base model must be Apache-2.0. In 2026 that means FLUX.1 [schnell] or a 4B FLUX.2 klein model.
• The klein split is by size, not by quantization: across 13 klein repositories, every 4B variant is Apache-2.0 and ungated, and every 9B variant is other and gated.
• Flux LoRA file sizes vary roughly 28× in the wild, which tells you about rank, not quality.

This is a reading of the published licence text, not legal advice. The wording quoted below is what the documents actually say on the date given; check them yourself before shipping anything commercial.

The Flux family in 2026

Before the licence question makes sense, the roster has to be current — and it has changed a great deal since the FLUX.1 launch that most tutorials still describe.

Two things stand out. FLUX.2 [dev] is 32.22 billion parameters — nearly three times FLUX.1 [dev], which puts full fine-tuning far out of consumer reach and makes LoRA the only realistic option. And klein 4B is the only model here you can download without accepting a gate, which is not a coincidence: it is also the only one with no restriction to accept.

Is a LoRA a “Derivative”? The licence answers directly

This is the question every Flux LoRA tutorial skips, and it is not ambiguous. The FLUX [dev] Non-Commercial License defines the term like this:

“Derivative” means any (i) modified version of the FLUX [dev] Model (including but not limited to any customized or fine-tuned version thereof), (ii) work based on the FLUX [dev] Model, or (iii) any other derivative work thereof.

A LoRA is a fine-tuned version. It is therefore a Derivative, and the grant that follows is limited:

You may only access, use, Distribute, or create Derivatives of the FLUX [dev] Model or Derivatives for Non-Commercial Purposes.

And so there is no doubt about scope, the licence adds: “Any restrictions set forth herein regarding the FLUX [dev] Model also apply to any Derivative you create or that are created on your behalf.”

“Non-Commercial Purpose” is defined tightly, and the definition is worth reading in full because it is narrower than people assume — it applies “only so far as you do not receive any direct or indirect payment”, and it explicitly carves out “(a) for revenue-generating activity, (b) in direct interactions with or that has impact on end users, or (c) to train, fine tune, or distill other models for commercial use”. Clause (b) is the one that catches people: a free tool that your customers touch is still outside the grant.

The images are treated differently — but read both clauses

Here is the distinction that actually matters day to day, and it runs the other way:

For the avoidance of doubt, Outputs are not considered Derivatives under this License. … We claim no ownership rights in and to the Outputs. … You may use Output for any purpose (including for commercial purposes), except as expressly prohibited herein.

This clause is narrower than it first looks, and the two halves of the licence pull against each other. The Outputs clause governs the artefacts: images you generated during use the licence already permits are yours outright, and you may use them commercially later. It does not widen the grant covering the model. That grant is limited to Non-Commercial Purposes, which apply “only so far as you do not receive any direct or indirect payment arising from the use of the FLUX [dev] Model” and expressly exclude use “(a) for revenue-generating activity” and “(b) in direct interactions with or that has impact on end users”.

So generating images for a paying client, or running a [dev] model inside a product, is outside the grant even though the resulting images would themselves be unencumbered — the payment arises from using the model. That is precisely the gap Black Forest Labs sells a commercial licence to close. The one express prohibition on Outputs is separate: you “may not use the Output to train, fine-tune, or distill a model that is competitive with” the FLUX models.

There are now two licence documents, and one supersedes the other

Worth knowing if you are citing terms: the original FLUX.1 [dev] Non-Commercial License v1.1.1 lives in the flux repository and enumerates the FLUX.1 models by name. The newer FLUX [dev] Non-Commercial License v2.0, in the flux2 repository, is written for the whole family — it covers “FLUX.x [dev], where “.x” denotes the FLUX model version number”, listing FLUX.1 [dev] through FLUX.2 [dev]. The Derivative and Output clauses are materially the same in both. A third document, the FLUX Non-Commercial License v2.1, sits alongside it for models flagged by an attached licence notice.

The klein split is the practical route to a commercial LoRA

FLUX.2 klein ships in a lot of variants — base and instruct, fp8, nvfp4, kv — and it would be easy to assume the licence follows the quantization. It does not. Across all 13 klein repositories the pattern has no exceptions:

• Every 4B variant is Apache-2.0, and ungated: klein-4B, klein-base-4B, klein-4b-fp8, klein-base-4b-fp8, klein-4b-nvfp4, klein-base-4b-nvfp4.
• Every 9B variant is other, and gated: klein-9B, klein-base-9B, klein-9b-fp8, klein-9b-kv, klein-9b-kv-fp8, klein-9b-nvfp4, klein-base-9b-nvfp4.

The split is on parameter count alone. For anyone who needs to distribute or sell the adapter, that makes klein 4B the default choice, and it is a 3.88B model rather than a 32B one, so it trains on far less hardware. The older route — FLUX.1 [schnell], also Apache-2.0 at 11.89B — is still open, and remains the better pick if you specifically want FLUX.1 behaviour.

For what the training run itself will cost in memory, our LoRA training VRAM guide works the arithmetic through for Flux, SDXL and SD 3.5, and the LoRA training VRAM calculator computes it for your own settings. This page deliberately does not repeat those figures.

Why Flux LoRA downloads differ by 28×

Two popular public Flux LoRAs, measured from their actual files:

Nothing about the larger file makes it better. LoRA size is set almost entirely by rank and by how many modules were targeted; a rank-128 adapter over every attention block is simply a bigger file than a rank-16 adapter over a few. What the size does predict is VRAM at load time and how much you are storing per style. If you want the underlying concept rather than the Flux specifics, we cover checkpoints, LoRAs and VAEs separately.

Choosing, in one paragraph

If the LoRA is for you and the images are the product, FLUX.1 [dev] or FLUX.2 [dev] are fine and are the strongest models on offer — you keep the images outright. If the LoRA itself will be sold, shipped, or placed in front of users, start from an Apache-2.0 base: klein 4B for new work, FLUX.1 [schnell] if you need the older behaviour. And if you have already trained on a [dev] model and now want to commercialise it, the licence points you at one route only — requesting a commercial licence from Black Forest Labs, which the text says “may be subject to a fee, royalty or other revenue share”.

FAQ

Can I sell images made with a FLUX.1 [dev] LoRA?

Images you generated during permitted non-commercial use are yours: Outputs are not Derivatives, Black Forest Labs claims no ownership, and you may use them for any purpose including commercial ones. But that does not make paid generation itself permitted — if the payment arises from operating the model, you are outside the Non-Commercial grant and need a commercial licence. The one express prohibition on Outputs is using them to train a competitive model.

Can I sell or distribute the LoRA file itself?

Not if it was trained on a [dev] model. The LoRA is a Derivative, and Derivatives are limited to Non-Commercial Purposes. Train on FLUX.1 [schnell] or a 4B klein model instead, both Apache-2.0. For the mechanics of loading and weighting an adapter once you have one, see our Stable Diffusion LoRA guide.

Which Flux model is easiest to train a LoRA on?

FLUX.2 klein 4B, at 3.88B parameters — roughly a third the size of FLUX.1 [dev] and about a ninth of FLUX.2 [dev] at 32.22B. It is also the only one you can download without a gate.

Is FLUX.2 klein 9B non-commercial?

Yes. Every 9B klein repository carries the other licence tag and is gated, while every 4B one is Apache-2.0 and ungated. The split follows parameter count, not quantization format.

Sources

• FLUX.1 [dev] Non-Commercial License v1.1.1 (github.com/black-forest-labs/flux)
• FLUX [dev] Non-Commercial License v2.0 and FLUX Non-Commercial License v2.1 (github.com/black-forest-labs/flux2)
• FLUX.1 [schnell], Apache-2.0 (huggingface.co)
• FLUX.2 klein 4B, Apache-2.0 and ungated (huggingface.co)
• Parameter counts, licence tags, gating status and file sizes read from the Hugging Face API on 26 September 2026

Updated 26 September 2026: licence texts, parameter counts, gating status and LoRA file sizes all read from primary sources on this date. Model licences change — re-check before relying on any of this commercially.