Soulkyn

Most AI girlfriend sites are a chat window with an image button bolted on. Soulkyn is a whole lot more ambitious: this is a sprawling toolkit with an active community.

I signed up, worked through the UK age check (sadly now a reality!), then burned through the free trial with a maid character called Janine. I came away pretty impressed.

Chat: Gwem vs GLM 5.2

Soulkyn's text chat is one of the major draws, and it currently runs on two models. Standard chats, including the free trial and Premium, use Gwem, which Soulkyn describes as its fast in-house model.

Deluxe and Deluxe Plus move you onto GLM 5.2 automatically.

Soulkyn makes some pretty big claims here: it says GLM 5.2 is a 753B+ parameter model running on its own GPU clusters, that it benches "between Opus-class and the very top frontier assistants" for roleplay, and that self-hosting is why Deluxe chat has no caps or cooldowns.

I don't want to dwell too much on specific models, because chances are - they'll be several months behind by the time you read this - but it's rare for an AI girlfriend app to be anywhere near the frontier.

What I can judge is the Gwem output from my initial session. It was actually pretty good. Replies started streaming in 3 to 6 seconds and finished in 6 to 11, which feels snappy. Long enough to feel like there's somebody behind a keyboard, not so long that I feel like I'm getting screwed out of priority tokens.

The prose is in classic roleplay style, with actions in asterisks and dialogue in between, and you'll likely know straight up whether this is going to "cut it" for you.

We traded messages back and forth and it was clear that the AI follows instructions well.

When I asked for a short answer I got one line ("A hot bath, a glass of that white wine… and someone skilled waiting on your every need"), when I asked to say hi in French in one sentence I got exactly that, and 12 times 12 came back as "Cent quarante-quatre, that's 144, monsieur", correct and in character.

The model held scene continuity too, which is always a primary concern... especially as the page begins to scroll...

It's not flawless. My virtual boo leaned on the same notebook gag twice in a few messages, the kind of repetition that can snowball in long threads. Every reply carries the same asterisk-action framing too, so if you prefer plain dialogue... you'll have to say as much (and hope for the best). I had mixed results with my forays into dirty chat...

Memory and context

Soulkyn uses two systems over every thread. Per-message recall searches the whole thread each turn and pulls back the most relevant past moments, and then you've got these rolling summaries which condense the conversation as it goes (seems to be after around 50 messages).

On top of that you get forced memories: a text box for capturing anything that you want to be factored into every reply. Best to keep this short.

Character creator

Want to build your own AI girlfriend?

The creation page offers five routes that all end in the same editor: a manual builder, Create with AI from a plain-text description, Spin the Wheel (a gacha pull), Breed (combine two Kyns into a third with mutations) and Import, either from another service or from an image.

The manual wizard runs four steps covering identity, appearance, personality, then name and advanced settings, and it's pretty strict with what it wants: my first attempt was blocked until I gave it at least three identity traits.

Every edit creates a new version with unlimited history and rollback, which I haven't really seen done this cleanly elsewhere. Kyns also carry an S/T/M/E/X content rating with 0 to 10 scores for things like sexual content and dark themes, plus a level and XP counter. On free you only get one custom and one AI-created character, and the AI-assisted, gacha, breeding and image-import routes are all Premium features behind the paywall.

Images

Soulkyn is not just a wall of text.

The image generation has seven models split between two families. I believe the SD family (SD Anime Original, Anime Alpha and Anime Hyper) is Stable Diffusion based. The other natural-language family (SK-Alpha, ZIT Sk-All, Sk-Photo and Sk-Real) takes plain sentences and swaps most of those dials for a Force Style toggle (with the familiar leading choice off Realistic vs Anime that we've seen repeated pretty much everywhere.)

SD models take 550 characters of prompt, up to three LoRAs and batches of 2, 4 or 6. Natural-language models take 1,400 characters, up to four LoRAs and batches of 1, 2 or 4. There are 63 LoRAs in 23 categories, face reference on the natural-language models for keeping a character consistent, img2img from a Soulkyn image link on the ZIT models (not SK-Alpha), and a bunch of saveable presets.

The output is impressive, and it's getting better with every new model release.

I was taken aback by natural skin with visible freckles, convincing green eyes, clean long hair and plausible office lighting - but more impressively, her face stayed recognisable across six different shots, outfits and rooms, which suggests the face consistency tools are finally beginning to stick without the classic drift between iterations.

The weaknesses are kinda familiar: a flat, slightly overexposed indoor look, exaggerated proportions as the default house style, and hands mostly kept out of frame.... wise but sensible!. Every image carries a small soulkyn.com watermark.

The anime side is more mixed. A neon office gallery kept the character's outfit, hair and lanyard consistent across six images, but the background signage was the usual AI gibberish text, and the poses repeated heavily. A fantasy RPG set had atmospheric candlelit halls with crowds of background figures, though those crowds turned soft and smeared up close.

Explore thumbnails were polished and colourful, with a sharp, cel-shaded look on the anime side and heavy airbrushed skin on the semi-realistic ones. Overall I would put the best realistic work near the top of the companion-site pack, and the anime work as decent... but nothing spectacular.