Capabilities and Limits

Scope: these limits are enforced by Lithune. When you use the service through OpenRouter, OpenRouter may apply its own limits as well. Prices are shown on OpenRouter.

Model

Model	Qwen3.8 27B (AWQ INT4 · Text and image input)
OpenRouter model ID	`qwen/qwen3.8-27b`
Context window	262,144 tokens (prompt and output together)
Maximum output	131,072 tokens
Default output limit	16,384 tokens when `max_tokens` is not set
KV cache	FP8
API	OpenAI-compatible chat completions, streaming supported
Hardware location	Türkiye

Input

Text and images are the only accepted inputs. PDFs, audio, video and other files are not accepted; render PDF pages to images first.
Images are sent as image_url content parts, either as an HTTPS URL or as a base64 data URL.
Formats: PNG, JPEG, WEBP, GIF. The format is detected from the image bytes, not from the declared type.
Up to 64 images per request, counted across all messages.
Up to 64 MiB per image and 128 MiB for all images in a request, counted after base64 decoding.
Up to 67,108,864 pixels per source image.
Each image is resized for the model to between 65,536 and 16,777,216 pixels, with both sides a multiple of 32 and the aspect ratio kept as closely as possible. Aspect ratios above 200:1 are rejected. The image's token cost follows this processed size.
image_url.detail is accepted but ignored.
The whole request body may be up to 100 MiB.

Image URLs

Only HTTPS URLs on public internet addresses are fetched.
URLs may be up to 4,096 bytes long, and up to 3 redirects are followed.
Each image must be downloaded within 15 seconds, and all images of a request within 60 seconds.
Downloads identify themselves with the user agent Lithune-Media-Acquirer/1.

Thinking

Thinking is on by default.
Choose the effort with reasoning.effort or reasoning_effort. Accepted values: none, low, medium, xhigh. none turns thinking off; reasoning.enabled: false does the same.
The reasoning is returned separately from the answer. Hide it with include_reasoning: false or reasoning.exclude: true.
Thinking budgets (reasoning.max_tokens) are not supported.

Sampling parameters you leave unset take these defaults:

Parameter	Thinking on	Thinking off
`temperature`	1.0	0.7
`top_p`	0.95	0.80
`top_k`	20	20
`min_p`	0.0	0.0
`presence_penalty`	0.0	1.5
`repetition_penalty`	1.0	1.0

Parameters

Parameter	Accepted values
`temperature`	0 to 2
`top_p`	0.01 to 1
`top_k`	0 or more
`min_p`	0 to 1
`frequency_penalty`	-2 to 2
`presence_penalty`	-2 to 2
`repetition_penalty`	0.01 to 2
`logit_bias`	-100 to 100 per token
`stop`	Up to 4 sequences
`max_tokens`, `max_completion_tokens`	1 to 131,072
`seed`	Any integer
`tools`, `tool_choice`	Function calling
`response_format`	JSON mode and JSON Schema structured outputs

Values outside these ranges are rejected with an error instead of being clamped. n greater than 1, logprobs, prompt_logprobs and echo are not supported.

Prompt caching

Prompt prefixes are cached automatically; no request option is needed. The number of prompt tokens served from the cache is reported in usage.prompt_tokens_details.cached_tokens. Cached blocks stay only in volatile memory; see the Data Policy.