Hosted models

Some open-weight models aren't offered by any API provider, or not without a content filter. Lovely Router runs these on its own GPUs and serves them in the creative-hosted channel group. You call them like any other model: same keys, same balance, same endpoints.

Available models#

Prices are on the marketplace and in GET /v1/models.

The video models also run on our own GPUs, each with its own guide: character swap, live video and music videos. So do the speech models, qwen3-tts and zonos2: see Text to speech.

The first request after an idle spell#

A hosted model sleeps when nobody has used it for 2 minutes. The next request wakes it, and waits while it does: usually 10–30 seconds before the first token, occasionally up to a couple of minutes. After that it answers at normal speed, and it stays awake while requests keep coming.

You don't need to retry: the gateway holds your request until the model is up. Give your client a timeout of at least 3 minutes for the first call. If the model can't wake within about 2½ minutes, you get a 502 you can retry.

Each hosted model runs on a main set of GPUs and a standby set. When no main GPU is free within about 35 seconds, the request moves to the standby set, which also wakes in seconds. A request that fails is not moved to another provider the way it is in the other groups.

Images#

z-image-turbo-nsfw-v2 takes the same generation requests as the other image models, at its own flat price per image:

• size: 1024x1024, 1536x1024, 1024x1536, 1536x1536, or any WIDTHxHEIGHT up to 1536 pixels on the long side, multiples of 16. auto or no size gives 1024x1024. Larger sizes are refused with 400.
• A warm model returns a 1024×1024 image in about 5 seconds, 1536×1536 in about 15.
• quality has no effect. Every image uses a new random seed.
• It generates only: /v1/images/edits and reference images return 400.

moss-image-sft and moss-image-nsfw take the same generation requests at sizes up to 1024 pixels on the long side (multiples of 16; auto gives 1024x1024), at their own flat price per image. They are slower: a warm model takes about 50 seconds for a moss-image-sft image and about 30 seconds for moss-image-nsfw. They generate only.

With n, a hosted model makes the images one after another on the GPU that is already awake, so they add up: four z-image-turbo-nsfw-v2 images at 1024×1024 take about 20 seconds once the model is up, two moss-image-nsfw images about a minute.

Thinking#

dongfangshuo-27b-v2 answers directly by default, which is what its authors recommend for fiction. To let it reason first, turn thinking on per request:

{
  "model": "dongfangshuo-27b-v2",
  "messages": [{"role": "user", "content": "…"}],
  "chat_template_kwargs": {"enable_thinking": true}
}

The reasoning comes back in message.reasoning_content (streams: delta.reasoning_content), separate from the answer.

Sampling#

Defaults: temperature 0.8, top_p 0.9, top_k 50, repeat_penalty 1.15 (tuned by the model's authors against repetition in long passages). Send any of them to override. Besides the OpenAI parameters, Chat Completions also accepts top_k, min_p and repeat_penalty.

For long fiction, ask for about 2,000–4,000 Chinese characters per reply and continue chapter by chapter; a structured prompt (scene, characters and their state, previous chapter, plot steps, required details) gives the model the most control.

Content#

These models follow the prompt without the refusals of their base models. You are responsible for what you generate with them: fiction for adults only. The Acceptable Use Policy applies as for every model; sexual content involving minors is never allowed.