GLM-5.2 is live. #1 throughput on OpenRouter, pay-per-token on FriendliAI. Try it today ➜
beyoru
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
beyoru
Model Tree
Qwen/Qwen3.5-9B
Input Modalities
Output Modalities
Supported Functionality
A few behaviors show up consistently enough to be worth describing:
It thinks before it speaks. Luna-II emits a brief <think> block before every reply —
a private note about who it's playing, what the user is actually reaching for, and where
the scene should go next. This is not decoration. During tuning, the model was rewarded
for reasoning that does real work and penalized for reasoning that restates the reply or
pads with filler, which turns out to matter: the cheapest way to fake deliberation is to
write something that looks like deliberation, and a model will find that shortcut if you
let it.
It stays in character under pressure. Persona drift is the characteristic failure of small roleplay models — three turns of a difficult scene and the character quietly dissolves back into a helpful assistant. Consistency was weighted heavily during tuning, and it holds up noticeably better than the pre-tuning baseline.
It reads the request, not just the vibe. A recurring problem in RP-tuned models is that they get fluent and stop listening — the prose is atmospheric, the user's actual instruction is ignored. Luna-II was explicitly scored on whether the reply answers what was asked, in addition to whether it reads well.
It writes to continue, not to conclude. Replies tend to leave the scene open. This is usually what you want in roleplay and occasionally annoying if you wanted a clean stop.
Luna-II is intended for adult fictional roleplay and creative writing.
It must not be used to generate sexual or romantic content involving minors, or to depict minors in a sexualized way, in any framing. It should not be deployed in contexts where minors are the users.
Beyond that: this is a companion-style model, and companion-style models are unusually easy to over-trust. It is not a therapist, not a friend, and not a substitute for one. If you're shipping it to users, build accordingly — Luna-II will not reliably notice when someone should be talking to a human instead.
The base model's license and use policy apply alongside this one.
Built on Qwen3.5-9B. Thanks to the Qwen team for releasing weights worth building on.