Qwen¶

Qwen extends the OpenAI-compatible endpoint with QwenModelConfig. Use it when your server needs Qwen-specific settings or returns tool calls embedded in XML.

Connection¶

import intuit;

auto ep = new Qwen("http://localhost:1234");
auto cfg = cast(QwenModelConfig)ep.config("qwen3-30b-a3b");
cfg.enableThinking = false;
cfg.topK = 40;

The constructor accepts a base url, optional API key, and display name (default "Qwen"). Authentication and request paths are inherited from OpenAI. Discovery and explicit configuration both create QwenModelConfig instances.

Replace the URL and model ID with those accepted by your server. Qwen-specific fields aren't supported by every OpenAI-compatible deployment.

Completion Settings¶

Field	Behavior
`enableThinking`	Defaults to `true`. Setting it to `false` sends `enable_thinking: false` only when `chatTemplateKwargs` is null.
`chatTemplateKwargs`	Sends `chat_template_kwargs` when non-null. Takes precedence over the `enableThinking` shortcut.
`topK`	Sends `top_k` when nonnegative.
`thinkingBudget`	Sends `thinking_budget` when nonnegative.
`mmProcessorKwargs`	Sends `mm_processor_kwargs` when non-null.

The base completion settings are applied first. Qwen-specific settings are added afterward and can replace the corresponding fields in params.

Embedding Settings¶

encodingFormat defaults to "float" and is omitted at that value. dimensions is sent only when positive. These settings affect embedding requests, not completions.

The embedding parser expects numeric arrays. Setting a non-float wire encoding doesn't add a decoder for that format, so keep the default unless your server still returns numeric vectors.

XML Tool Calls¶

Native OpenAI-style tool_calls take precedence. When no native calls are present, the parser looks for supported <tool_call> blocks in the response text.

It accepts Qwen3-Coder function/parameter tags and Hermes-style JSON-in-XML forms. Extracted calls populate Choice.toolCalls, receive synthetic IDs, and are removed from Choice.text. Surrounding prose is preserved. Unrecognized blocks aren't guaranteed to produce calls.

Extraction doesn't authorize execution. Use the shared Tools workflow to register functions, evaluate policy, and handle pending calls.