There is no single model called reve-v1. The string maps to at least four different things, and one of them has nothing to do with model versions at all: api.reve.com/v1/ is the API path version, and it serves Reve’s current models rather than any first-generation one.
If you are picking a model today, the answer is Reve 2.1, released 9 July 2026. Everything with a 1.x in the name is at least a generation behind, and in some cases two.
The Four Things That Name Could Mean
| What you might mean | What it actually is | Status |
|---|---|---|
| Reve Image 1.0 | The original model, codenamed Halfmoon | Superseded twice |
| Reve v1.1 / v1.1 Fast | A distinct listed model with a low-latency variant | Previous generation |
| Reve Image 1.5 | Another separately listed 1.x model | Previous generation |
api.reve.com/v1/ |
The API path version, not a model | Current, serves latest models |
There is a fifth wrinkle. Picsart’s model registry lists Reve under the identifier reve:reve@1, labelled simply as Reve 1, which is a platform-level version number rather than Reve’s own model naming.
So if you found reve-v1 in someone’s code, check what it is actually pointing at before assuming it is the first-generation model. In the official API, the version parameter is separate from the path, and leaving it unset gives you the latest model rather than an old one.
The Actual Version History
Reve AI, Inc. is a Palo Alto company founded by researchers who came out of Google Brain and NVIDIA.
It surfaced in March 2025 with an unannounced preview model nicknamed Halfmoon, which climbed to the top of the Artificial Analysis Image Arena ahead of Midjourney v6.1 and Google Imagen 3 before anyone knew who had built it. Halfmoon was later revealed as Reve Image 1.0.
What made 1.0 unusual was the training approach. Rather than learning from captions, it was trained on detailed data structures describing composition, spatial relationships, and style. That decision matters more in hindsight than it did at the time, because the whole current architecture grew out of it.
Reve 2.0 arrived in June 2026 and re-architected everything around layouts and native 4K output. Reve 2.1 followed roughly a month later, on 9 July 2026, with reportedly around three times the parameters of its predecessor, sharper prompt understanding, and noticeably better foreign-script text rendering.
Two major releases in five weeks is a fast cadence. Treat any Reve article written before July 2026 as describing a different product.
The Architecture Is the Point
Most image models go straight from text to pixels. Reve’s pitch is that this is the wrong shape for anything you need to revise.
The company’s own framing is that most generators are still in a fireworks phase: you load the tube, light it, and hope. If a hand comes out wrong or a sign in your image reads RSETAURANT, your only move is to re-roll and hope the next attempt fixes it without breaking something else.
Reve splits generation into two steps. The model first plans the image as a structured layout, then renders from it. Because that intermediate representation persists, every element in the finished image stays addressable. You can change one region and re-render just that region, without the rest degrading through repeated regeneration.
The company summarises the product split as the images being the code and the website being the editor. That is a genuinely different mental model from prompt-and-pray, and it is the reason Reve keeps appearing in production pipelines rather than just in art threads.
What Reve 2.1 Actually Does
- Native 4K at 16 megapixels, not upscaled from a smaller render
- Region-level editing that preserves sharp edges and texture across repeated edits
- Multilingual text rendering, including dense legible typography in foreign scripts
- Up to eight reference images, addressed individually in the prompt by index
- Edit and remix in one request, where the prompt decides how much of the reference survives
That last one is a neat design choice. Asking to change the wall colour in a reference keeps the scene and repaints the wall. Asking for the handbag from that reference on a model walking a sunlit street keeps the product’s identity and rebuilds everything around it. Same endpoint, different instruction.
Text rendering deserves the emphasis it gets. Legible typography inside generated images has been one of the field’s persistent failures, and it is the difference between a usable ad mockup and a nice picture you cannot ship.
Where to Access It, and What It Costs
Reve is available directly through its own API and through a long list of intermediaries, including ComfyUI partner nodes, Picsart, Runware, and several aggregator APIs.
Pricing varies by route and this is where care is needed. Runware lists Reve 2.1 generation starting at $0.20 per 4K image. Other coverage cites edits at around a cent per operation. Those are not contradictory, since generating a fresh 4K frame and nudging one region are different amounts of work, but reseller markups sit on top of whichever number you find.
Price from the vendor you will actually bill through, and if you are comparing against other models, our AI image generator comparison covers the wider field on the same terms.
Rights and Openness
Reve-v1 is closed source. Despite early community speculation that the weights might be released, they have not been, as of the most recent reporting. The team did fork a ComfyUI repository and became an official ComfyUI partner, but that is workflow integration rather than open weights.
On commercial use, reporting indicates Reve grants full commercial rights on generated images including on its free tier, which is more generous than several competitors. Verify that on Reve’s own terms before building a campaign on it, since these policies change and the company has moved domains during its short history.
Worth remembering that a platform granting you commercial permission is a separate question from whether you hold enforceable copyright in the output. We covered that distinction in detail when looking at AI content ownership and copyright, and it applies here identically.
Who Should Use It
Strong fit: anyone whose work involves revising an image rather than generating and discarding. Product photography variants, ad mockups with real copy in them, and any pipeline where a client comes back asking to change one element. The lossless editing story is the reason to be here.
Also strong: teams needing text inside images in more than one script. Few competitors handle this well.
Look elsewhere if you want open weights you can self-host, or if your work is stylistic exploration rather than precise composition. For that, our roundup of the best AI art generators covers models tuned more for aesthetic range than for control.
Frequently Asked Questions
Is reve-v1 the same as Reve 2.1?
No. Reve 2.1 is the current model, released 9 July 2026. A v1 reference points either to a 1.x-generation model or to the API path version, which is not a model identifier at all.
What happened to Halfmoon?
Halfmoon was the codename for Reve Image 1.0, launched March 2025. It was superseded by Reve 2.0 in June 2026 and Reve 2.1 in July 2026. The name no longer refers to anything current.
Can I run Reve locally?
No. The model is closed source and the weights have not been published, despite early speculation that they might be.
The Practical Call
If you are writing code, do not pin to a 1.x model on the assumption that it is stable and current. Leave the version parameter unset to get the latest, or pin explicitly to 2.1 if you need reproducible output.
And if you were evaluating Reve on the basis of what the original Halfmoon model did in 2025, evaluate it again. The architecture changed twice since then, and the thing worth judging now is not whether it makes attractive images but whether the layout-first editing survives the fifth revision on a real client job. That is the claim it is actually making.
Published: September 10, 2026
