Open and closed AI models: what the weights decide

Give ten practices the same massing model and the same AI renderer, and the images come back related. Not identical, related. The same low afternoon sun, the same immaculate glazing, the same couple crossing the plaza with a dog. You can push it around by prompting harder, but the family resemblance survives, because the prompt did not put it there. Training did, months before anyone in your office opened the tool.

So the useful question is not which renderer wins this quarter's comparison. It is whether you can change what the thing prefers. That depends almost entirely on one property of the model you are using: whether its weights are open or closed.

What you are actually renting

A model is a very large set of numbers, called weights, fixed at the end of its training and unchanged afterward. Generating an image means running your prompt through those numbers. Everything the model knows about brick, glare, handrail proportion, and how a car park reads at dusk lives in there, and nowhere else.

Closed models keep the weights on the maker's servers. You send text and reference images to an app or an API, and pixels come back. Midjourney works this way, as does Google's Nano Banana Pro (Gemini 3 Pro Image), generally available since June 2026 at around thirteen cents per image at 1K or 2K. You rent the operation without ever holding the thing that performs it.

Open-weight models publish the file. FLUX.2 [dev], released by Black Forest Labs in November 2025, is a 32-billion-parameter model whose weights sit on Hugging Face for anyone to download and run on their own machine. The FLUX.2 [klein] family that followed in January 2026 is smaller and faster, and its 4-billion-parameter version carries an Apache 2.0 licence. Stable Diffusion 3.5 has been downloadable since October 2024. In each case the file is yours to keep, run offline, and modify.

FIG. 01 - Where the weights live. Placeholder. Artwork to follow: two-column diagram, hairline weight only. Left column closed, the prompt crossing a boundary line and pixels returning, weights drawn inside the vendor's boundary. Right column open, weights inside the office boundary, nothing crossing the line.

Why this decides more than image quality

Three consequences reach the way a practice actually works.

House style. Studios have a way of drawing, and it is one of the few things a client recognises before they can say why. A closed model can be steered image by image with references, style codes, and careful prompts, but it cannot be taught once and thereafter know. Every project restarts the negotiation, and so does every new person in the office.

What leaves the office. A closed model means uploading geometry, viewports, and reference material to somebody else's servers: competition schemes, unbuilt work, occasionally a site before it is announced. That may be perfectly acceptable, but it is worth checking against the confidentiality clause you signed, which almost certainly predates the practice of sending a client's massing to a rendering API. A self-hosted open-weight model sends nothing anywhere.

Version stability. Midjourney's V7 was the default from June 2025 until June 2026, when V8.1 replaced it, and V8.2 became default in July 2026. Each version renders differently. A look you built on a hosted model moves when the vendor moves, on the vendor's schedule, while a downloaded checkpoint is frozen and will render the same next year as it does today. Both behaviours are useful. Only one is under your control.

Training is the specification you never wrote

Nobody wrote rules into these models. A model was shown an enormous number of images paired with text, and its weights were adjusted over and over until it could produce plausible images from a description. What comes out the far end is not knowledge in any tidy sense. It is a set of defaults inherited directly from what the model was shown, so when the training images skew toward marketing renders and property photography, that is the taste you have hired.

The closest thing in a practice is the standard specification. Nobody re-argues the substrate under a floor finish on every job; the spec decides it, quietly, a few thousand times a year, and the resulting building is shaped by decisions no one made consciously on that project. A model's weights work the same way. You can override the spec on a drawing and you can override the model in a prompt, but you are fighting the default each time rather than changing it. The analogy fails at one point, and the failure is the part worth keeping: a specification is legible. You can open it, read the clause, and see exactly why the floor is what it is. Nobody can open a set of weights and find the clause that makes the model love glass. The only way to change what it prefers is to show it more of something else.

FIG. 02 - Training as inherited default. Placeholder. Artwork to follow: a horizontal band of flat thumbnails, the same generic render repeating with small variations, one frame at the right breaking the pattern and outlined in terracotta.

Showing it more of something else has a name: fine-tuning. You take a finished model and continue its training on a small, deliberate set of images until its defaults shift toward yours. Fifty of your own drawings, consistently captioned, will do more for how a model renders your work than any prompt you will ever write. The light, cheap version of fine-tuning is called LoRA, it is what practices actually use, and it is the subject of the next article.

What matters here is the dependency. Fine-tuning writes into the numbers, so it requires holding the numbers. That is why open versus closed is not a philosophical preference but a fork in what your practice can build.

Open weights is not open source

The two get used interchangeably and they are not the same thing. Open weights means the file is downloadable. It does not mean the training data is published, and for the image models architects actually use, open or closed, it is not. You cannot audit what a model learned from, which is a real limit when a client asks where an image came from.

Nor does downloadable mean unrestricted. FLUX.2 [dev] weights are free to obtain under a non-commercial licence; commercial self-hosting, along with the right to fine-tune and use LoRAs commercially, is sold separately by Black Forest Labs in tiers with monthly image allowances. Stable Diffusion 3.5 is free to use commercially under Stability's community licence only while the organisation's total annual revenue stays under one million dollars, counted across the whole business rather than the part that makes images. Read the licence before the technology, not after.

FIG. 03 - What each licence lets you do. Placeholder. Artwork to follow: drawing-sheet table, hairline rules only. Rows: FLUX.2 [dev], FLUX.2 [klein] 4B, Stable Diffusion 3.5, Nano Banana Pro, Midjourney. Columns: weights downloadable, self-host, fine-tune, commercial use. Ticks and dashes only, terracotta reserved for the restriction column.

Then there is the cost that never appears on an invoice. Running these models locally needs a capable GPU, and someone in the office has to own the setup and the training runs. Black Forest Labs says an optimised implementation of FLUX.2 [dev] runs on consumer RTX cards, and the klein models generate in under a second on the same hardware, so the barrier is lower than it was two years ago. It is not zero.

What the closed side is still better at

The frontier closed models remain ahead on things architects care about: legible text inside an image, factual grounding, and reliability without anyone maintaining anything. For a practice that renders occasionally and has no appetite for infrastructure, that is the correct trade, and paying by the image is an honest way to buy it.

Closed vendors do sell customisation of a kind. Midjourney's personalisation profiles learn your preferences from the images you rate, and Google offers custom style tuning for Imagen to approved users through an access request. Both are useful. Neither gives you a file. That tuning lives on the vendor's side, cannot be moved to another tool, and does not necessarily survive the next model version.

The question to ask before you build a house style

Start by finding out which model sits underneath the tool your office uses most. Many rendering tools are interfaces over models they did not build, and the marketing rarely leads with it, but the answer is usually one support page away. Once you know the model, you know whether the weights are open, and therefore whether the door is open.

Then ask the question that decides whether the effort is worth spending: if we teach this thing to draw the way we draw, who holds the result, and where does it live? Those ten related renders are not a prompting failure to be fixed with better adjectives. They are what training looks like from the outside, and only training answers them. The next article goes through the door: an afternoon, a folder of your own images, and a small file that carries your studio's habits.