← Back to blog
AiAbout 6 min read

Flux 3 and Imagen 4 Land on Replicate, and the Image Shelf Gets Crowded

Published Oct 2, 2026
Flux 3 and Imagen 4 Land on Replicate, and the Image Shelf Gets Crowded

On 1 October, Black Forest Labs and Google both put new image models on Replicate. Black Forest Labs released Flux 3 for text-to-image and editing, with support for up to ten reference images. Google put four models up: Imagen 4 fast for speed and cost, Imagen 4 ultra for detail, Imagen 3 fast for price-sensitive work, and Imagen 3 for top-quality output. Call them through the same API you already use.

The headline reads like a model launch. It is closer to a distribution announcement. Every one of these models already existed somewhere. What changed is that a developer with a Replicate token can now reach them without opening an account with Black Forest Labs or wiring up a Google Cloud project.

What ten reference images actually buys you

The number worth pausing on is Flux 3's ten reference images. A year ago, image editing meant one input picture and a text instruction. Ten references changes what the model can be asked to do. You can hand it a product shot, a texture from a second image, a colour palette from a third, and a face from a fourth, then ask for a composition that respects all of them.

That is the capability the whole market has been racing toward. Nano Banana Pro blends up to fourteen images and five people. GPT Image 2.5 Sunburst leads the Artificial Analysis editing board by making one region change without disturbing the rest. Seedream 5.0 reaches for live web search when an edit needs a real-world reference. The competitive line has moved from how good the first image looks to how well the model takes direction on the tenth edit.

A dense wall of small image thumbnails of varied scenes with one tile glowing brighter and floating out from the grid

For anyone building a product on top of these tools, reference-image count is a more useful spec than benchmark scores. It tells you how much control you can hand a user. A model that takes one reference suits a text-to-image button. A model that takes ten suits a design tool where the user assembles branded material from a folder of approved assets.

Google's four models are really two tiers

Google's lineup looks like four releases. Look at the names and it is two models split by cost. Imagen 4 fast and Imagen 4 ultra run the same generation, one tuned for cheapness and one for detail. Same for Imagen 3 fast and Imagen 3. The fast variants are not worse models in a meaningful sense. They are the same model with a lower compute budget, sold at a lower price.

That is a sensible way to sell image generation, and it is also a hint about where the cost sits. When the only difference between a fast and an ultra tier is how long the model is allowed to think, the model itself has stopped being the product. Compute has become the product, and the model is the thing that spends it.

For a team deciding what to run in production, that changes the calculus. The question is not which of the four is best. It is which quality level your use case actually requires, because paying for ultra on a use case that only needs fast is a way to burn money without improving anything a user will notice.

What a platform listing is actually worth

It is worth being precise about why a model landing on Replicate is news. Replicate rents out hosted inference. A developer signs up once, gets a token, and calls any model in the catalog through the same interface. For Black Forest Labs and Google, listing means reaching developers who will never open a separate account for one model, and paying a platform for the reach.

This is the deal open model makers have been forced into. Publish weights and the technically curious can run them, at no cost to anyone and with no revenue to the lab. List on a hosted platform and the audience widens to people who want an endpoint and a monthly bill, but a middleman takes a slice and the model sits beside a dozen competitors in a picker.

The alternative is to sell access directly, which is what a closed model does and what most labs cannot do at scale. So the hosted platform becomes the neutral ground where open and closed models compete for the same developer, judged on the same page. A small lab with good weights can now show up next to a Google release, which is either healthy competition or a commoditisation race depending on which side of it you are on.

For a buyer, that has one practical consequence. The model you pick today is reachable through an interface you already know, which makes switching costs low and experimenting cheap. It also means the model you pick is probably not special, because the same shelf is open to everyone.

The rest of the day

The same day produced a wide spread of smaller releases. Cloudflare put Clef, an image-text-to-text model, on Hugging Face. ISTA-DASLab published a GGUF build of Qwen3.8 Flash, which matters mostly to people running models locally. Vercel's AI Gateway added a decision model called Laya that routes support requests and agent workflows with probability scores, free through the end of October, alongside Microsoft audio models and search tools from Browserbase. Google added real-time visual assistance to Gemini Live, built with blind and low-vision users.

None of these will change an image pipeline on their own. Together they show where the platform vendors are pushing. Replicate, Vercel, and Hugging Face are competing to be the place where a developer picks a model, and the way they compete is by making the catalog wider and the billing simpler. The model labs benefit, because distribution through a platform is how a smaller lab reaches developers who will never sign a separate contract.

That is the trade every open model maker faces. Publish weights and let the community run them, which costs nothing and reaches the technically curious. Or list on a hosted platform and reach the developers who want an endpoint and a bill, which reaches more people and hands a platform a cut. Flux 3 is on Replicate, which suggests Black Forest Labs wants the second audience badly enough to share.

What it means if you are choosing

The practical effect of all this is that the shelf is crowded and the labels are getting harder to read. Four Imagen models, a Flux release, and a long tail of community builds all compete for the same slot in a pipeline. The differences that matter are narrower than the marketing suggests.

Three things are worth checking before you commit. How many reference images the model takes, because that determines the ceiling on control. Whether the licence permits commercial use, which is where Apache-2.0 and non-commercial builds split. And whether the model is available somewhere that fits how you deploy, since a great model behind a platform you cannot use is a model you cannot use.

None of that is glamorous, and none of it will be in a launch post. It is also the part that decides whether a model works in a product six months after the announcement has scrolled past.

Related articles