# q38 — free Qwen3.8-27B endpoint (no key)

Qwen3.8-27B, the best open model that runs on a consumer machine (Sept 2026).
Used as the lower bar in evals. Thai and English both work.

## Quick (GET, plain-text answer)
https://q38.wker.dev/?q=YOUR+QUESTION
Optional: system=… (instructions), max=… (output tokens, up to 8000),
effort=none|low|medium|high (thinking; default none), format=json (full JSON response).

## Full (POST JSON, OpenAI chat-completions style)
POST https://q38.wker.dev/
{"messages":[{"role":"user","content":"Explain compound interest in 3 lines."}],
 "max_completion_tokens": 300, "reasoning_effort": "none", "temperature": 0}
Shortcut: {"prompt":"…","system":"…"}. Also accepted: top_p, seed, response_format, tools, tool_choice, stream.
Response: {"model":"qwen/qwen3.8-27b","provider":"<host>","choices":[…],"usage":{…}}
"provider" is the host that served the call (cheapest bf16/fp8 host); log it with "model" for evals.

## OpenAI SDK / tools (base URL)
base_url = https://q38.wker.dev/v1   api_key = anything (ignored)   model = anything (always qwen3.8-27b)

## Limits
Output up to 8000 tokens (thinking counts), input up to 200000 characters, fair-use rate limits.
Do not send private or customer data: text passes through Cloudflare, OpenRouter and the host.
Sibling endpoints: https://luna.wker.dev (GPT-6 Luna), https://jev.wker.dev (yes/no or pick-one decisions).