A model's parameter count sets the local hardware bill

The process wall

Before a maker runs a large image or video model on their own machine rather than through someone else's API, the first constraint they meet is not the software but the graphics card's memory. A model's weights have to fit into VRAM before it can generate anything, and that figure is fixed by the model, not negotiable by the user, which is why the hardware step comes before the creative one in any local workflow.

What the documents show

Black Forest Labs' model card states FLUX.1 [dev] is a 12 billion parameter rectified flow transformer released under a non-commercial licence, and its own reference code includes a step to offload the model to CPU specifically to save VRAM when a card lacks enough GPU power to hold it. ComfyUI's system requirements documentation, current as retrieved on 16 September 2026, lists NVIDIA, AMD, Intel Arc and Apple Silicon as supported hardware for running such models locally through its node-based interface, without stating a single minimum VRAM figure. NVIDIA's own product page for its current flagship consumer card states the RTX 5090 carries 32 GB of GDDR7 memory — the ceiling a maker buying the best available consumer hardware works against.

Authorship and limitation

The mechanism is arithmetic, not opinion, and it is worth doing rather than taking on faith: loading a model in the reduced-precision format its own example code uses spends roughly two bytes per parameter, so twelve billion parameters occupy on the order of twenty-four gigabytes for the weights alone, before the activations a generation pass also needs — an editorial calculation drawn from the stated parameter count, not a figure either document states directly. That is close enough to a 32GB flagship card's ceiling that the model card's CPU-offload instruction is not a convenience feature; it is the documented workaround for a machine that cannot otherwise hold the model, at the cost of the slower transfer offloading involves.

What survives

A maker who checks a model's stated parameter count against a card's stated memory before buying anything can predict whether they will need offloading, a smaller model variant, or a rented API, rather than discovering it mid-generation. Editorially, VRAM is the practical budget line for local generative work in a way clock speed or a marketed performance figure is not, because a model that does not fit does not run at all.

• Check a target model's stated parameter count and loading precision before estimating memory needs.
• Read a local tool's system requirements page for which GPU vendors and offload options it supports.
• Compare that estimate against a specific card's stated memory configuration, not a whole product line's headline.

None of these documents publishes a single minimum-VRAM number for a given creative task, which means the planning step is still the maker's own arithmetic, done once, against figures each vendor does state plainly.