FLUX.1 dev vs schnell

The common assumption is that schnell is a cut-down dev, and the numbers say otherwise: their transformers are 23.78GB and 23.80GB, they share the same 9.77GB T5-XXL text encoder and the same VAE. Hardware requirements are effectively identical. What actually separates them is the licence — schnell is Apache 2.0 and dev is non-commercial — and the step count, since schnell is distilled to produce an image in one to four steps while dev expects a conventional multi-step run with guidance.

FLUX.1 [dev] vs FLUX.1 [schnell] at a glance

Peak VRAM at 1024x1024, batch 1 — the two models side by side

Computed by the same function as the image generation calculator, so these figures cannot drift from it. Activation memory is an approximation — plan with headroom.

The VRAM difference is a rounding error

We read the component sizes from the safetensors files in each model's repository and summed them per component. The transformers differ by 0.02GB. The text encoder is the same T5-XXL in both, at 9.77GB. The VAE is identical.

So any statement of the form 'use schnell if you have less VRAM' is wrong. Whatever offload strategy and precision you need for one, you need for the other. If dev does not fit on your card at FP8, schnell will not either.

The T5-XXL encoder is the interesting number in both. At 9.77GB it is a large fraction of the total, and it only has to be resident while the prompt is being encoded. Offloading it to the CPU frees close to ten gigabytes for the price of a couple of seconds per prompt, which is the single most effective thing you can do to fit either model on a smaller card.

Distillation is what schnell actually is

schnell is guidance-distilled. Rather than running the usual many-step denoising loop with classifier-free guidance, it has been trained so that a very small number of steps produces a finished image — typically one to four. That is where the name comes from.

The practical effect is on time, not memory. A one-to-four-step generation completes far quicker than a twenty-to-fifty-step one on the same hardware, which makes schnell noticeably more pleasant to iterate with on a card where each step is slow.

dev keeps the conventional sampling loop and guidance control, and is generally regarded as producing higher fidelity and responding better to detailed prompts and negative conditioning. You are trading iteration speed and licence freedom against that.

The licence is the decision

schnell is Apache 2.0. You can use the outputs commercially, build it into a product and distribute derivatives, with the usual Apache obligations and nothing more.

dev ships under a non-commercial licence. It is fine for personal work, research and experimentation, and it is not fine for anything that ends up in a commercial product — including client work and images used to promote a business. Black Forest Labs offers separate commercial terms for dev; if that is your situation, that is the conversation to have rather than hoping the restriction is unenforced.

So the practical rule: if the output is going anywhere near a product or a client, use schnell. If it is personal or experimental and you want maximum fidelity, use dev. Hardware does not enter into it.

Work it out for your own setup

Common questions

Does schnell need less VRAM than dev?

No. Their transformers differ by 0.02GB and they share the same text encoder and VAE. Whatever precision and offload strategy fits one will fit the other.

Can I use FLUX.1 [dev] for client work?

Not under the standard licence, which is non-commercial. Client work is commercial use. Black Forest Labs offers separate commercial terms for dev, so if you need dev specifically that is the route; otherwise schnell's Apache 2.0 licence removes the question.

Will Flux run on 8GB of VRAM?

Not by our own numbers. The cheapest configuration the calculator produces for either model is NF4 with sequential CPU offload, and that peaks at about 9.94GB at 1024x1024 — above an 8GB card. NF4 with the text encoder on CPU is about 10.11GB. In practice 12GB is the realistic floor for Flux, and FP8 with the encoder offloaded, at about 14.87GB, is the first configuration that is comfortable rather than desperate.

Is there a version of dev distilled for fewer steps?

Community LoRAs exist that reduce dev's step count. They do not change the licence, which is the constraint that matters if your use is commercial.