What AI model does SpicyChat use?

Several, and its server says which. In our signed-in test on October 21, 2026, every chat reply came back with an engine field. The default model answered as NextDayAI/G4-26B-A4B-Zeta-int4, SpicyChat's own model. Picking Qwen3 235B returned Qwen/Qwen3-235B-A22B-Instruct-2507, Lyra 12B returned Sao10K/MN-12B-Lyra-v4, Codex 24B returned Gryphe/Codex-24B-Small-3.2 and Stheno returned a Stheno v3.2 8B build. Its top plan lists DeepSeek V4 Pro, Kimi K2, Glam 5.1 and MiniMax Text 01 by name, but we couldn't test those.

What image model does DreamGF use?

Z-Image Turbo. Every picture DreamGF made for our account carried its own generation recipe inside the image file, and all of them load a quantized Z-Image Turbo checkpoint with a Qwen3-4B text encoder. DreamGF doesn't say which model writes its chat replies.

What models does Xotic use for video?

Xotic's own code credits outside engines: its talking videos say "Powered by OmniHuman", its V3 animations say "Powered by Oh.xyz", and its newest NOVA engine is marked "v5-wan3" with a WAN3_ENABLED switch. Its chat models are in-house names: Ivy, Zara, Selene and Scarlet.

Are SpicyChat and GirlfriendGPT the same company?

Both sites name the same operator in their footers: NextDay AI, with addresses in Montreal, Delaware and Limassol. Both card charges started with "NDAI", and the same client code lists both brands. GirlfriendGPT's own model menu only uses in-house names.

What AI Models Do SpicyChat, GirlfriendGPT, DreamGF, Xotic and Nomi Use?

Q: What AI model does Nomi use?

Nomi only uses its own names. Its chat models are Solstice (Stable), Cambrian 6 (Beta) and Mosaic (Legacy); pictures come from its version 5 image system with two processes called Riva and Lago. The only outside AI company in its code is ElevenLabs, as an optional bring-your-own-key voice.

Most AI girlfriend apps give their models pet names and leave it there. We wanted to know what’s underneath, so on October 21 we did what we did for Ourdream, Candy, Secrets and Kindroid: we captured each site, exported its code, and recorded what our own signed-in browser received while we chatted and made pictures. No private APIs, no tricks, nothing a curious user’s browser doesn’t already get.

The short answer: SpicyChat is the only one that tells you, reply by reply. DreamGF’s pictures confess in their own files. Xotic’s code credits three outside video engines. GirlfriendGPT and Nomi say nothing beyond their own names.

What we found

App	Chat	Pictures	Video	How sure
SpicyChat	Its own 26B model by default; Qwen3 235B, Lyra 12B, Codex 24B, Stheno 8B and more on request	Z-Image, FLUX.2 klein, Qwen Image (named in the picker)	None	Model names in the server’s own replies
GirlfriendGPT	In-house names only (Gem 2.6, Quinn III, Blaze XL…)	Seedream and Qwen in code	Yes, model not named	Code and app menus
DreamGF	Not disclosed	Z-Image Turbo	Not named	Recipe inside every picture
Xotic	In-house names (Ivy, Zara, Selene, Scarlet)	ComfyUI workflows; Seedream via fal in code	NOVA (“v5-wan3”), OmniHuman, Oh.xyz	Code and on-screen credits
Nomi	In-house names (Solstice, Cambrian 6, Mosaic)	Its own “Riva” and “Lago”	V3, model not named	Code and app menus

SpicyChat: the server names the model

This was the surprise. When SpicyChat answers you, the response that reaches your browser includes an engine field naming the model that wrote the reply. We switched between twelve models in its own menu on a True Supporter account and sent a message with each. Some of what came back:

Default answered as NextDayAI/G4-26B-A4B-Zeta-int4, a model filed under the name of SpicyChat’s own company. “Zeta G4 26b”, sold as a separate option, came back as the same engine.
Qwen3 235B 2507 answered as Qwen/Qwen3-235B-A22B-Instruct-2507, Alibaba’s open model. Its “Wildcard Mode” came back as the same checkpoint.
Lyra 12B V4 answered as Sao10K/MN-12B-Lyra-v4_fp8, a community roleplay fine-tune.
Codex 24B answered as Gryphe/Codex-24B-Small-3.2, and Squelching Fantasies 8B as Mawdistical/Squelching-Fantasies-8B, both community fine-tunes.
Magnum 12B was odd: one reply came back labelled with the default 26B engine, another with anthracite-org/magnum-v4-12b. We don’t know why.

The “I’m All In” plan lists more famous names, including DeepSeek V4 Pro, Kimi K2, Glam 5.1, MiniMax Text 01 and MiMo V2.6 Pro. We saw those only in the app’s model list, not in a reply, because they needed a plan we didn’t buy.

For pictures, the avatar maker’s own tiles say Z-Image, FLUX.2 klein and Qwen Image. The code also has an error message for image moderation by Sightengine. One more thing we enjoyed: asked which model it was, the character told us it was “a large language model, trained by Google”, in the same reply whose engine field said otherwise. Never trust a chatbot’s answer about itself.

GirlfriendGPT: same operator, quieter about it

GirlfriendGPT’s footer names the same company as SpicyChat’s: NextDay AI. Its model menu, though, uses only in-house names, each with a size label: Gem 2.6 (26B, the default), Ember Lite (8B), Lyra (12B), GGPT XL (132B), Quinn III (235B) and Blaze XL (355B). The stream for our chat named gem-2.6 and nothing more. Several of those size labels match SpicyChat’s lineup (a 26B default, a 235B, a 12B Lyra, a 132B XL), which is interesting but not proof that they’re the same models.

The code is more talkative about pictures and voice: its image list includes Seedream and Qwen (shown to users as “Character Lite”), and its voice settings name Inworld and Cartesia. We didn’t see either voice service in action, so treat those as configured, not confirmed.

DreamGF: the pictures carry their own recipe

DreamGF names nothing on screen, but its pictures do. Every image it made for our account was a PNG with the full generation recipe tucked inside the file: a quantized Z-Image Turbo checkpoint, a Qwen3-4B text encoder and Z-Image’s own decoder, eight sampling steps. All 16 files we checked said the same. (Qwen3-4B there is the part that reads the prompt, not the chatbot.) Z-Image Turbo is also what powers Ourdream’s “Vivid” style, so two of the apps we’ve tested draw with the same open model.

What writes DreamGF’s chat replies, we couldn’t find in the code, the network traffic or the files. The company behind it is DreamAI SRL in Bucharest, per its footer.

Xotic: credits on the video buttons

Xotic’s chat models are in-house: Ivy (free), Zara, Selene and Scarlet. Its pictures come from ComfyUI workflows behind four engines it calls Noir, Classic, Apex and Lumen, and the code has a fifth, hidden engine wired to Seedream via fal.

Video is where it credits other people. The talking-video option says “Powered by OmniHuman · Lip-synced to character voice”. The V3 animation says “Powered by Oh.xyz”. And its newest engine, NOVA, is tagged v5-wan3 in the code, next to a switch called WAN3_ENABLED. That points to the Wan video model family, though Xotic itself only shows the name NOVA.

Nomi: everything in-house

Nomi gives away the least. Its chat models are Solstice (Stable), Cambrian 6 (Beta) and Mosaic (Legacy), with Aurora and Odyssey still in the code. Pictures come from its version 5 image system, with two processes it calls Riva and Lago. Video is “V3”. The only outside AI company in the code is ElevenLabs, and only as an option to plug in your own voice account; Nomi says its built-in voices are “all in-house Nomi voices”.

Does any of this matter?

Less than the marketing suggests. The best conversation in our tests came from Xotic’s in-house Ivy and from Secrets, which names nothing at all. SpicyChat’s openness is still worth something: you can see when you’re on the free default and when you’re on Qwen3, and you can judge the upgrade for yourself.

What we couldn’t see: the models behind locked plans, anything that runs only on the apps’ servers, and every app’s chat model except SpicyChat’s. Apps change models often; this is what their sites sent on October 21, 2026.

More on each app: SpicyChat, GirlfriendGPT, DreamGF, Xotic, Nomi.