nexusflow
Online
API Reference

Parameter Matrix

Parameters are listed according to the backend's actual forwarding and protocol conversion logic. Text models support all three protocols (OpenAI, Anthropic, Responses); non-text models use the image, audio, embedding, or async task endpoints based on their capabilities.

OpenAI Chat Completions

Parameter
Type/Mapping
Status
Description
model
string
Required
Model ID. Text, reasoning, multimodal, coding, and specialized models support Chat Completions.
messages
array
Required
Array of conversation messages, passed in order: system, user, assistant, tool.
messages[].role
string
Required
system / user / assistant / tool. tool messages return function execution results.
messages[].content
string | array
Required
Plain text can be passed as a string; multimodal input uses an array of content blocks.
messages[].content[].type
string
Multimodal
Stable examples are text / image_url; extended content blocks like video and input_audio must be tested per model.
messages[].content[].text
string
Multimodal
Text when type=text.
messages[].content[].image_url.url
string
Multimodal
Image URL or data URL; requires a model with vision support.
stream
boolean
Optional
Enable SSE streaming. Recommended for long text, reasoning models, and interactive scenarios.
stream_options.include_usage
boolean
Optional
Returns usage at the end of a streaming response. Recommended for billing, statistics, or smoke checks.
temperature
number
Optional
Sampling temperature. Typically 0 to 2; higher is more random.
top_p
number
Optional
Nucleus sampling threshold. Avoid adjusting it heavily together with temperature.
max_tokens
integer
Optional
Maximum output tokens, cannot exceed the model maxOutput.
stop
string | string[]
Optional
Stop sequences; output ends when matched.
presence_penalty
number
Optional
Presence penalty, typically -2 to 2, encourages new topics.
frequency_penalty
number
Optional
Frequency penalty, typically -2 to 2, reduces repetition.
tools
array
Optional
Array of function calling definitions. Only models that support function calling reliably return tool_calls.
tools[].type
string
Tool
Always function.
tools[].function.name
string
Tool
Function name. Use letters, digits, and underscores.
tools[].function.description
string
Tool
Description of the function, affecting how accurately the model selects tools.
tools[].function.parameters
object
Tool
JSON Schema describing the function parameters.
tool_choice
string | object
Optional
Stably supports auto / none, or specify {type:'function', function:{name}}. Forcing tools is not recommended for thinking mode models.
response_format
object
Optional
Output format control. Common values are {"type":"text"} or {"type":"json_object"}.
enable_thinking
boolean
Optional
Thinking mode toggle. Can be turned off only for verified hybrid thinking models; thinking-only models ignore false and keep returning reasoning_content.
thinking_budget
integer
Optional
Caps the thinking token budget, forwarded by model ID prefix (qwen3.7- / qwen3.6- / qwen3.5- / qwen3-).
preserve_thinking
boolean
Optional
Forwards reasoning_content from prior messages back to the model. Supported by qwen3.7-max, qwen3.6-max-preview, qwen3.6-plus, kimi-k2.6.
enable_search
boolean
Optional
Web search, supported by Qwen text models (not the VL / math series).
search_options
object
Optional
Web search configuration, used together with enable_search.
enable_context_caching
boolean
Optional
Enable context cache. Repeated prompt prefixes are cached automatically; hits are billed at 0.1x the input price. Supported by Qwen and GLM series.
seed
integer
Optional
Random seed, forwarded for Qwen text models.
top_k
integer
Optional
Top-K sampling, forwarded for Qwen text models.
logprobs
boolean
Optional
Return log probabilities, forwarded for Qwen text models.
repetition_penalty
number
Optional
Repetition penalty, forwarded for Qwen text models.
parallel_tool_calls
boolean
Optional
Parallel function calling, supported by Qwen, DeepSeek, GLM, and Anthropic models.

Currently Unsupported Fields

The table below lists fields not yet stably forwarded by the public Chat endpoint; do not rely on them in production code.

Parameter
Type/Mapping
Status
Description
max_completion_tokens
integer
Not forwarded
Use the currently stable max_tokens instead.

Thinking Mode Support

This lists the tested behavior of the NexusFlow live OpenAI Chat endpoint. Support varies with upstream model versions, so production code should configure behavior explicitly per model ID.

Parameter
Type/Mapping
Status
Description
qwen3.7-max
Hybrid thinking
Supports true / false
Thinking on by default; true returns reasoning_content, false does not. Supports thinking_budget and preserve_thinking.
qwen3.5-flash
Hybrid thinking
Supports true / false
Verified live: true returns reasoning_content, false does not.
qwen3-max
Hybrid thinking
Supports true / false
Verified live: true returns reasoning_content, false does not.
qwq-plus
Thinking only
false cannot disable
Verified live: both true/false return reasoning_content.
qwen-math-plus
Not treated as a thinking toggle
Do not send
Verified live: neither true/false returns reasoning_content.
deepseek-r1
Thinking only
false cannot disable
Verified live: both true/false return reasoning_content.
deepseek-v3.2
Hybrid thinking
Supports true / false
Verified live: true returns reasoning_content, false does not.
deepseek-v4-pro
Hybrid thinking
Supports true / false
Verified live: true returns reasoning_content, false does not.
glm-5.2
Hybrid thinking
Supports true / false
Flagship for long-horizon tasks, 1M context. Thinking on by default; true returns reasoning_content with up to 128K chain-of-thought, false does not. Supports thinking_budget.
glm-5.1
Hybrid thinking
Supports true / false
Verified live: true returns reasoning_content, false does not.

Anthropic Messages Mapping

Parameter
Type/Mapping
Status
model
model
Model ID, mapped to the OpenAI model.
system
messages[0].role=system
System prompt. Supports a string or text blocks.
messages
messages
user / assistant messages are converted to OpenAI messages.
messages[].content[].text
messages[].content
Text block. Plain 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
Returns the function execution result.
max_tokens
max_tokens
Maximum output tokens.
temperature
temperature
Sampling temperature.
top_p
top_p
Nucleus sampling.
stop_sequences
stop
Array of stop sequences.
stream
stream
Enable the 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.

Responses API Mapping

Parameter
Type/Mapping
Status
model
model
Model name, e.g. qwen3.7-plus.
input
messages
Plain text or an array of messages (supports role: user/assistant/system/developer).
instructions
system message
System instructions, inserted at the start of the context.
previous_response_id
—
Previous response ID for multi-turn conversations (valid for 7 days).
stream
stream
Whether to enable streaming output.
store
—
Whether to store the response (default true); if false, it cannot be referenced via previous_response_id.
tools
tools
Tool list: web_search, web_extractor, code_interpreter, function, etc.
tool_choice
tool_choice
Tool selection strategy: auto / none / required.
temperature
temperature
Sampling temperature.
top_p
top_p
Nucleus sampling.
max_output_tokens
max_tokens
Maximum output tokens.
enable_thinking
enable_thinking
Whether to enable thinking mode.
reasoning
—
Thinking effort control, e.g. {effort: "high"}.

Response Fields

Parameter
Type/Mapping
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 delta.
choices[].delta.reasoning_content
Streaming thinking delta that reasoning models may return.
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 tip: for reasoning models, use stream=true and stream_options.include_usage=true; for hybrid thinking models in low-cost, low-latency scenarios, explicitly pass enable_thinking=false. See more examples in the Chat Completions API and the Responses API.