API Reference

Parameters Matrix

This page lists parameters based on backend actual forwarding and protocol conversion logic. Text-type models support OpenAI, Anthropic, and Gemini protocols; non-text models use image, audio, embedding, or async task interfaces based on model capabilities.

OpenAI Chat Completions

Parameter

Type/Map

Status

Description

model

string

Required

Model ID. Text, reasoning, multimodal, coding, and professional models support Chat Completions.

messages

array

Required

Chat message array, passed in sequential order: system, user, assistant, tool.

messages[].role

string

Required

system / user / assistant / tool. The tool role is used for passing back tool execution results.

messages[].content

string | array

Required

Text can be passed directly as a string; multimodal input uses a content block array.

messages[].content[].type

string

Multimodal

Common values include text / image_url; video, input_audio, and other Bailian extension content blocks depend on the specific model.

messages[].content[].text

string

Multimodal

Text content when type=text.

messages[].content[].image_url.url

string

Multimodal

Image URL or data URL; requires the model to support vision understanding.

stream

boolean

Optional

Enable SSE streaming output. Recommended for long text, reasoning models, and interactive scenarios.

stream_options.include_usage

boolean

Optional

Return usage in the final streaming response chunk. Recommended when billing, statistics, or smoke checks are needed.

temperature

number

Optional

Sampling temperature. Range typically 0 to 2; higher values produce more random output.

top_p

number

Optional

Top-p sampling threshold. It is recommended not to adjust both temperature and top_p significantly at the same time.

max_tokens

integer

Optional

Maximum output token count; cannot exceed the model's maximum output.

stop

string | string[]

Optional

Stop sequences; the model stops output after hitting one.

presence_penalty

number

Optional

Presence penalty, typical range -2 to 2, increases tendency for new topics.

frequency_penalty

number

Optional

Frequency penalty, typical range -2 to 2, reduces repetition in output.

tools

array

Optional

Function calling definition array. Only models that support tool calling will reliably return tool_calls.

tools[].type

string

Tool

Fixed as function.

tools[].function.name

string

Tool

Function name. Recommended to use letters, digits, and underscores.

tools[].function.description

string

Tool

Description of the function's purpose, affects the model's accuracy in selecting tools.

tools[].function.parameters

object

Tool

JSON Schema describing the function's input parameters.

tool_choice

string | object

Optional

Supports auto / none, or specifying {type:'function', function:{name}}. For thinking mode models, forcing specific tools is not recommended.

response_format

object

Optional

Output format control. Common values: {"type":"text"} or {"type":"json_object"}.

enable_thinking

boolean

Optional

Toggle thinking mode. Only verified hybrid thinking models can be disabled; pure thinking models will ignore false and continue returning reasoning_content.

thinking_budget

integer

Optional

Limit thinking token upper bound, passed through based on model ID prefix (qwen3.7- / qwen3.6- / qwen3.5- / qwen3-).

preserve_thinking

boolean

Optional

Pass historical reasoning_content in messages back to the model. Supports qwen3.7-max, qwen3.6-max-preview, qwen3.6-plus, kimi-k2.6.

enable_search

boolean

Optional

Web search, supported by Qwen (Tongyi) text-type models (not VL / math series).

search_options

object

Optional

Web search configuration, used together with enable_search.

enable_context_caching

boolean

Optional

Enable Context Caching. Repeated prompt prefixes are automatically cached; hits are billed at 0.1x input price. Supports Qwen (Tongyi), GLM series.

seed

integer

Optional

Random seed, supported by Qwen (Tongyi) text models for pass-through.

top_k

integer

Optional

Top-K sampling, supported by Qwen (Tongyi) text models for pass-through.

logprobs

boolean

Optional

Return log probabilities, supported by Qwen (Tongyi) text models for pass-through.

repetition_penalty

number

Optional

Repetition penalty, supported by Qwen (Tongyi) text models for pass-through.

parallel_tool_calls

boolean

Optional

Parallel tool calling, supported by Qwen (Tongyi), DeepSeek, GLM, Anthropic models.

Not Yet Supported Fields

The table below lists fields that are not yet reliably forwarded through the public Chat endpoint. Production code should not depend on these.

Parameter

Type/Map

Status

Description

max_completion_tokens

integer

Not yet forwarded

Please use the currently supported max_tokens instead.

Thinking Mode Support

This section lists the actual behavior of NexusFlow's online OpenAI Chat endpoint. Support may change with upstream model versions; production code should use explicit configuration based on model ID.

Parameter

Type/Map

Status

Description

qwen3.7-max

hybrid thinking

Supports true / false

Thinking enabled by default; true returns reasoning_content; false does not. Supports thinking_budget and preserve_thinking.

qwen3.5-flash

hybrid thinking

Supports true / false

Verified: true returns reasoning_content; false does not.

qwen3-max

hybrid thinking

Supports true / false

Verified: true returns reasoning_content; false does not.

qwq-plus

pure thinking

false cannot be disabled

Verified: true/false both return reasoning_content.

qwen-math-plus

not handled as thinking toggle

Do not pass

Verified: true/false do not yet return reasoning_content.

deepseek-r1

pure thinking

false cannot be disabled

Verified: true/false both return reasoning_content.

deepseek-v3.2

hybrid thinking

Supports true / false

Verified: true returns reasoning_content; false does not.

deepseek-v4-pro

hybrid thinking

Supports true / false

Verified: true returns reasoning_content; false does not.

glm-5.1

hybrid thinking

Supports true / false

Verified: true returns reasoning_content; false does not.

Anthropic Messages Mapping

Parameter

Type/Map

Status

model

model

Model ID, mapped to OpenAI model.

system

messages[0].role=system

System prompt. Supports string or text blocks.

messages

messages

user / assistant messages are converted to OpenAI messages.

messages[].content[].text

messages[].content

Text block. Pure text blocks are merged into a string.

messages[].content[].image

image_url

Supports url or base64 source, converted to OpenAI image_url.

messages[].content[].tool_use

assistant.tool_calls

Assistant tool call result.

messages[].content[].tool_result

role=tool

Tool execution result passed back.

max_tokens

max_tokens

Maximum output tokens.

temperature

temperature

Sampling temperature.

top_p

top_p

Nucleus sampling.

stop_sequences

stop

Stop sequence array.

stream

stream

Enable Anthropic SSE event stream.

tools

tools

Anthropic tools are converted to OpenAI function tools.

tool_choice

tool_choice

auto / none / any / tool is converted to OpenAI tool_choice.

Gemini GenerateContent Mapping

Parameter

Type/Map

Status

contents

messages

Message array. String contents are also wrapped into user text messages.

contents[].role

messages[].role

user maps to user, model maps to assistant.

contents[].parts[].text

content text

Text content.

contents[].parts[].inlineData

image_url data URL

Base64 image content, converted to image_url.

contents[].parts[].fileData

image_url

File URL, converted to image_url.

contents[].parts[].functionCall

assistant.tool_calls

Model function calling.

contents[].parts[].functionResponse

role=tool

Tool execution result.

systemInstruction

system message

System prompt, supports string or parts.

generationConfig.temperature

temperature

Sampling temperature.

generationConfig.topP

top_p

Nucleus sampling.

generationConfig.maxOutputTokens

max_tokens

Maximum output tokens.

generationConfig.stopSequences

stop

Stop sequence array.

tools[].functionDeclarations

tools

Function declarations, converted to OpenAI function tools.

toolConfig.functionCallingConfig.mode

tool_choice

AUTO / ANY / NONE map to auto / required / none respectively; some upstream models may not decline required.

streamGenerateContent

stream=true

Streaming interface. Use ?alt=sse for SSE-formatted responses.

Response Fields

Parameter

Type/Map

choices[].message.content

Non-streaming text output.

choices[].message.reasoning_content

Thinking content field that reasoning models may return.

choices[].message.tool_calls

Returned when the model requests a tool call.

choices[].delta.content

Streaming text increment.

choices[].delta.reasoning_content

Streaming thinking increment, may be returned by reasoning models.

choices[].finish_reason

stop / length / tool_calls / content_filter.

usage.prompt_tokens

Input tokens.

usage.completion_tokens

Output tokens.

usage.total_tokens

Total tokens.

usage.completion_tokens_details.reasoning_tokens

Reasoning tokens, returned by some models.

Production recommendation: Use stream=true and stream_options.include_usage=true for reasoning models; for hybrid thinking models in low-cost, low-latency scenarios, explicitly pass enable_thinking=false. More examples at Chat Completions API and Gemini Protocol.