> ## Documentation Index
> Fetch the complete documentation index at: https://evolink.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# GLM All-Model API - Chat Completions Reference
> - Call GLM series models through the OpenAI Chat Completions protocol, choosing the specific model with the `model` parameter
- Synchronous processing that returns the conversation content in real time
- **Text conversation**: single-turn or multi-turn contextual conversation; `glm-5.3-flash` additionally supports image input
- **System prompt**: customise the AI's role and behaviour through a `role=system` message
- **Deep thinking**: `thinking.type` controls the chain of thought and `reasoning_effort` adjusts the reasoning depth; the reasoning is returned in `reasoning_content`
- **Streaming**: SSE streaming responses are supported (`stream=true`)
- **Tool calling**: function calling and web search are supported (`web_search`, up to 128 tools)
- **Structured output**: JSON mode is enabled through `response_format`
**About streaming responses**: when `stream=true`, results come back over Server-Sent Events, each message formatted as `data: {JSON}` and the stream ending with `data: [DONE]`. Each chunk (`ChatCompletionChunk`) carries `id`, `created`, `model`, `choices` and optionally `usage` and `content_filter`; inside it, `choices[].delta` returns incremental `role` / `content` / `reasoning_content` / `tool_calls`, and `choices[].finish_reason` gives the termination reason in the final chunk.
**BaseURL**: The default BaseURL is `https://direct.evolink.ai`, which has better support for text models and long-lived connections. `https://api.evolink.ai` is the primary endpoint for multimodal services and serves as a fallback address for text models.
**Thinking control varies by model**: `glm-5.3` and `glm-5.3-flash` always think and cannot be turned off; for `reasoning_effort`, the three levels `low` / `high` / `max` take effect and the remaining levels are downgraded automatically to the nearest available one (`xhigh` → `max`, `medium` → `high`, `minimal` / `none` → `low`, which still thinks and is billed as output). `glm-5.2` can turn thinking off with `thinking.type: "disabled"` and supports more reasoning levels. See the `thinking` and `reasoning_effort` field descriptions for details.
**Image input**: supported only by `glm-5.3-flash`, through `image_url` content blocks inside `messages[].content[]`. Sending image blocks to any other model returns an error.
## OpenAPI
````yaml en/api-manual/language-series/glm/chat-completions/chat-completions-reference.json POST /v1/chat/completions
openapi: 3.1.0
info:
title: GLM All-Model API - Chat Completions Reference
description: >-
Full API reference for calling Zhipu GLM text models through the
OpenAI-compatible Chat Completions API.
**Models covered**: `glm-5.3`, `glm-5.3-flash`, `glm-5.2` (selected with the
`model` parameter)
**Shared capabilities**:
- 1M token context window, up to **131,072 tokens** (128K) of output, and at
least **1,024 tokens** recommended
- Deep thinking: `thinking` controls the chain of thought and
`reasoning_effort` adjusts the reasoning depth; the reasoning is returned in
`reasoning_content`
- Tool calling: function calling and web search (up to 128 tools)
- Streaming: SSE streaming responses
- Structured output: the `text` and `json_object` response formats
- Context cache: an implicit prefix cache that repeated requests with the
same prefix hit automatically, reported in
`usage.prompt_tokens_details.cached_tokens`
**Differences between models** (thinking control, image input) are described
in the `model`, `thinking` and `reasoning_effort` fields below.
license:
name: MIT
version: 1.0.0
servers:
- url: https://direct.evolink.ai
description: Production (recommended)
- url: https://api.evolink.ai
description: Alternative URL
security:
- bearerAuth: []
tags:
- name: Chat Completion
description: AI chat completion related endpoints
paths:
/v1/chat/completions:
post:
tags:
- Chat Completion
summary: GLM Chat Completions (All Models, OpenAI-Compatible)
description: >-
- Call GLM series models through the OpenAI Chat Completions protocol,
choosing the specific model with the `model` parameter
- Synchronous processing that returns the conversation content in real
time
- **Text conversation**: single-turn or multi-turn contextual
conversation; `glm-5.3-flash` additionally supports image input
- **System prompt**: customise the AI's role and behaviour through a
`role=system` message
- **Deep thinking**: `thinking.type` controls the chain of thought and
`reasoning_effort` adjusts the reasoning depth; the reasoning is
returned in `reasoning_content`
- **Streaming**: SSE streaming responses are supported (`stream=true`)
- **Tool calling**: function calling and web search are supported
(`web_search`, up to 128 tools)
- **Structured output**: JSON mode is enabled through `response_format`
**About streaming responses**: when `stream=true`, results come back
over Server-Sent Events, each message formatted as `data: {JSON}` and
the stream ending with `data: [DONE]`. Each chunk
(`ChatCompletionChunk`) carries `id`, `created`, `model`, `choices` and
optionally `usage` and `content_filter`; inside it, `choices[].delta`
returns incremental `role` / `content` / `reasoning_content` /
`tool_calls`, and `choices[].finish_reason` gives the termination reason
in the final chunk.
operationId: createChatCompletionGLM
requestBody:
required: true
content:
application/json:
schema:
$ref: '#/components/schemas/ChatCompletionRequest'
examples:
simple_text:
summary: Single-turn text conversation
value:
model: glm-5.3
messages:
- role: user
content: Please introduce yourself
multi_turn:
summary: Multi-turn conversation (context understanding)
value:
model: glm-5.3
messages:
- role: user
content: What is Python?
- role: assistant
content: Python is a high-level programming language...
- role: user
content: What are its advantages?
system_prompt:
summary: Using system prompts
value:
model: glm-5.3
messages:
- role: system
content: >-
You are a professional Python programming assistant.
Answer questions concisely.
- role: user
content: How do I read a file?
deep_thinking:
summary: Enable deep thinking and adjust reasoning intensity
value:
model: glm-5.3
messages:
- role: user
content: >-
A farmer needs to take a wolf, a sheep, and a cabbage
across a river, but can only carry one at a time. How
can he get them across safely?
thinking:
type: enabled
reasoning_effort: max
function_calling:
summary: Tool calling (Function Calling)
value:
model: glm-5.3
messages:
- role: user
content: What's the weather like in Beijing today?
tools:
- type: function
function:
name: get_weather
description: Query the real-time weather of a specified city
parameters:
type: object
properties:
city:
type: string
description: 'City name, for example: Beijing'
required:
- city
tool_choice: auto
web_search:
summary: Enable the web search tool
value:
model: glm-5.3
messages:
- role: user
content: >-
Help me find the latest artificial intelligence news
from the past week
tools:
- type: web_search
web_search:
enable: true
count: 10
search_recency_filter: oneWeek
description: >-
When the model decides it needs the web, it searches
automatically and the results are merged into the context for
reasoning. The retrieved results are billed as input tokens,
and the search service itself is billed per call (see the
pricing page); no search charge applies when the model does
not trigger a search.
json_mode:
summary: JSON structured output
value:
model: glm-5.3
messages:
- role: system
content: >-
Please output in JSON format, including the two fields
name and age.
- role: user
content: John, 28 years old this year
response_format:
type: json_object
streaming:
summary: Streaming output (SSE)
value:
model: glm-5.3
messages:
- role: user
content: Write a short poem about spring
stream: true
disable_thinking_glm52_only:
summary: Disable deep thinking (glm-5.2 only)
value:
model: glm-5.2
messages:
- role: user
content: Summarize the theory of relativity in one sentence.
thinking:
type: disabled
description: >-
Only `glm-5.2` can turn thinking off. Sending `thinking.type:
"disabled"` to `glm-5.3` or `glm-5.3-flash` returns an error —
use `reasoning_effort: "low"` instead.
low_effort:
summary: >-
Reduce thinking overhead (the alternative to disabling
thinking on the glm-5.3 series)
description: >-
The `glm-5.3` series cannot turn thinking off; use
`reasoning_effort: "low"` to bring the reasoning depth down to
its minimum.
value:
model: glm-5.3
messages:
- role: user
content: Explain HTTP in one sentence
thinking:
type: enabled
reasoning_effort: low
max_tokens: 1024
vision_flash:
summary: Image input (glm-5.3-flash only)
description: >-
`glm-5.3-flash` supports vision natively; images are passed in
through `image_url` content blocks, using either a public URL
or a Base64 data URL.
value:
model: glm-5.3-flash
messages:
- role: user
content:
- type: text
text: What is in this image?
- type: image_url
image_url:
url: https://example.com/photo.jpg
max_tokens: 1024
responses:
'200':
description: Chat completion generated successfully
content:
application/json:
schema:
$ref: '#/components/schemas/ChatCompletionResponse'
'400':
description: Invalid request parameters
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 400
message: Invalid request parameters
type: invalid_request_error
'401':
description: Unauthorized, invalid or expired token
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 401
message: Invalid or expired token
type: authentication_error
'402':
description: Insufficient quota, recharge required
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 402
message: Insufficient quota
type: insufficient_quota_error
fallback_suggestion: https://evolink.ai/dashboard/billing
'403':
description: Access denied
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 403
message: Access denied for this model
type: permission_error
param: model
'404':
description: Resource not found
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 404
message: Specified model not found
type: not_found_error
param: model
fallback_suggestion: glm-5.3
'429':
description: Rate limit exceeded
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 429
message: Rate limit exceeded
type: rate_limit_error
fallback_suggestion: retry after 60 seconds
'500':
description: Internal server error
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 500
message: Internal server error
type: internal_server_error
fallback_suggestion: try again later
'502':
description: Upstream service error
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 502
message: Upstream AI service unavailable
type: upstream_error
fallback_suggestion: try different model
'503':
description: Service temporarily unavailable
content:
application/json:
schema:
$ref: '#/components/schemas/ErrorResponse'
example:
error:
code: 503
message: Service temporarily unavailable
type: service_unavailable_error
fallback_suggestion: retry after 30 seconds
components:
schemas:
ChatCompletionRequest:
type: object
required:
- model
- messages
properties:
model:
type: string
description: >
Model to call:
| Model ID | Positioning | Thinking control | Image input |
|---|---|---|---|
| `glm-5.3` | Flagship model, with across-the-board gains on complex
software engineering and agent tasks and a large step up in coding;
1M context | Always thinks, cannot be turned off; only `low` /
`high` / `max` take effect for `reasoning_effort`, other levels are
downgraded automatically | Not supported |
| `glm-5.3-flash` | Lightweight multimodal model with a hybrid
sparse and linear attention architecture, extremely low cost and
native vision; 1M context | Same as `glm-5.3` | **Supported**, see
the `messages` field |
| `glm-5.2` | Previous-generation flagship, complex reasoning and
very long context; 1M context | Can be turned off with
`thinking.type: "disabled"`; `reasoning_effort` supports all 7
levels | Not supported |
enum:
- glm-5.3
- glm-5.3-flash
- glm-5.2
default: glm-5.3
example: glm-5.3
messages:
type: array
description: >-
The list of conversation messages, containing the full context of
the current conversation
Supports four roles: `system`, `user`, `assistant`, `tool`. Messages
of different roles have different field structures; please select
the corresponding role to view details. Must contain at least 1
message, and cannot consist solely of system or assistant messages.
items:
oneOf:
- $ref: '#/components/schemas/SystemMessage'
- $ref: '#/components/schemas/UserMessage'
- $ref: '#/components/schemas/AssistantRequestMessage'
- $ref: '#/components/schemas/ToolMessage'
discriminator:
propertyName: role
mapping:
system:
$ref: '#/components/schemas/SystemMessage'
user:
$ref: '#/components/schemas/UserMessage'
assistant:
$ref: '#/components/schemas/AssistantRequestMessage'
tool:
$ref: '#/components/schemas/ToolMessage'
minItems: 1
stream:
type: boolean
description: >-
Whether to enable streaming output mode
- `false`: The model generates the complete response and returns it
all at once (default), suitable for short text and batch processing
- `true`: Returns chunks in real time via Server-Sent Events (SSE),
suitable for chat and long text; returns `data: [DONE]` when the
stream ends
default: false
example: false
thinking:
type: object
description: Controls whether to enable the chain of thought (Chain of Thought)
properties:
type:
type: string
description: >-
Chain-of-thought switch
- `enabled`: turn on deep thinking (the default behaviour for
every model)
- `disabled`: turn off deep thinking, the model answers directly
**Only `glm-5.2` supports `disabled`.** `glm-5.3` and
`glm-5.3-flash` always think, and sending `disabled` returns an
error.
To lower thinking overhead on the `glm-5.3` series, use
`reasoning_effort: "low"` instead.
**Migrating from `glm-5.2`**: if your code hard-codes
`thinking.type: "disabled"`, you must change it to `"enabled"`
before switching to `glm-5.3` (adding `reasoning_effort: "low"`
if you want lower thinking overhead), otherwise the request
fails outright.
Note that the two fields behave **differently**: an unsupported
`reasoning_effort` level is downgraded automatically without an
error, whereas `thinking.type` is an explicit switch —
`disabled` always errors and is never silently rewritten. So
changing `reasoning_effort` from `none` to `low` on its own is
not enough; `thinking.type` has to be changed as well.
enum:
- enabled
- disabled
default: enabled
clear_thinking:
type: boolean
description: >-
Whether to clear the `reasoning_content` from previous
conversation turns
- `true` (default): Ignores/removes `reasoning_content` from
previous turns, using only non-reasoning content
(user/assistant-visible text, tool calls and results, etc.) as
context, which can reduce context length and cost
- `false`: Retains `reasoning_content` from previous turns and
provides it to the model along with the context (Preserved
Thinking); in this case, the historical `reasoning_content` must
be passed through in `messages` **completely, unmodified, and in
the original order**—missing, truncating, rewriting, or
reordering it will degrade results or prevent it from taking
effect
- Note: This parameter only affects historical thinking across
turns; it does not change whether thinking is produced in the
current turn
default: true
example: true
reasoning_effort:
type: string
description: >-
Controls how much the model reasons; takes effect only when
`thinking` is on, defaulting to `max`
**Supported values differ by model**:
| Value | `glm-5.3` / `glm-5.3-flash` | `glm-5.2` |
|---|---|---|
| `max` | Deep reasoning (default) | Deep reasoning |
| `high` | Enhanced reasoning | Enhanced reasoning |
| `low` | Light reasoning | Same as `high` |
| `xhigh` | Downgraded to `max` | Same as `max` |
| `medium` | Downgraded to `high` | Same as `high` |
| `minimal` | Downgraded to `low` (still thinks) | Skips thinking |
| `none` | Downgraded to `low` (still thinks) | Skips thinking |
The `glm-5.3` series always thinks, and only the three levels `low`
/ `high` / `max` genuinely take effect; the other four do not error
but are downgraded automatically to the nearest available level
(`xhigh` → `max`, `medium` → `high`, `minimal` / `none` → `low`).
**The `glm-5.3` series cannot turn thinking off.** Sending `minimal`
or `none` only drops it to the lowest level `low`; the model still
produces thinking tokens, **billed at the output rate**. If you send
these two levels to save money, note that this differs from
`glm-5.2` — on `glm-5.2` they genuinely skip thinking.
For complex tasks such as programming, `max` is recommended.
enum:
- max
- xhigh
- high
- medium
- low
- minimal
- none
default: max
example: max
do_sample:
type: boolean
description: >-
Whether to enable the sampling strategy
- `true` (default): Uses `temperature` / `top_p` for random
sampling, producing more varied output
- `false`: Always selects the highest-probability token (greedy
decoding), producing more deterministic output; in this case
`temperature` and `top_p` are ignored
For tasks requiring consistency and reproducibility (such as code
generation and translation), setting this to `false` is recommended
default: true
example: true
temperature:
type: number
format: float
description: >-
Sampling temperature, controlling the randomness and creativity of
the output
**Notes**:
- Value range: `[0.0, 1.0]`, limited to two decimal places
- Higher values (e.g. 0.8): more random and creative, suitable for
creative writing
- Lower values (e.g. 0.2): more stable and deterministic, suitable
for factual Q&A and code generation
- Default value: `1.0`
**Recommendation**: Do not adjust both `temperature` and `top_p` at
the same time
minimum: 0
maximum: 1
default: 1
example: 1
top_p:
type: number
format: float
description: >-
Nucleus Sampling parameter, an alternative to `temperature` sampling
**Notes**:
- Value range: `[0.01, 1.0]`, limited to two decimal places
- The model only considers candidate tokens whose cumulative
probability reaches `top_p`; for example, 0.1 means only the top 10%
probability tokens are considered
- Smaller values produce more focused and consistent output; larger
values increase diversity
- Default value: `0.95`
**Recommendation**: Do not adjust both `temperature` and `top_p` at
the same time
minimum: 0.01
maximum: 1
default: 0.95
example: 0.95
max_tokens:
type: integer
description: >-
Upper limit on the number of output tokens
**Note**:
- The GLM series supports up to **131,072 tokens** (128K) of output;
setting at least `1024` is recommended
- When `thinking` is on, chain-of-thought tokens also count towards
this limit
- If generation is cut off with `length`, try raising this value
minimum: 1
maximum: 131072
example: 1024
tools:
type: array
description: >-
The list of tools the model may call
**Note**:
- Function calling (`function`) and web search (`web_search`) are
supported
- Up to 128 functions
- Of these, `web_search` is **billed separately per call** when a
search actually happens; the other tools carry no extra charge
items:
oneOf:
- $ref: '#/components/schemas/FunctionTool'
- $ref: '#/components/schemas/WebSearchTool'
discriminator:
propertyName: type
mapping:
function:
$ref: '#/components/schemas/FunctionTool'
web_search:
$ref: '#/components/schemas/WebSearchTool'
maxItems: 128
tool_choice:
type: string
description: >-
Controls how the model selects which function to call
**Notes**: Only takes effect when the tool type is `function`;
defaults to and only supports `auto` (the model automatically
decides whether to call a tool)
enum:
- auto
default: auto
example: auto
stop:
type: array
description: >-
The list of stop words
**Notes**:
- When the generated text encounters a specified string, generation
stops immediately (the stop word itself is not included in the
returned text)
- Currently only a single stop word is supported, in the format
`["stop_word1"]`, for example `["Human:"]`
items:
type: string
maxItems: 4
example:
- 'Human:'
response_format:
type: object
description: >-
Specifies the model's response output format; defaults to `text`
**Notes**:
- `{ "type": "json_object" }` enables JSON mode, and the model
returns valid JSON-formatted data, suitable for scenarios such as
structured data extraction
- When using JSON mode, it is recommended to explicitly request JSON
output in the `system` or `user` message
required:
- type
properties:
type:
type: string
description: |-
Output format type
- `text`: Plain text output (default)
- `json_object`: JSON-formatted output
enum:
- text
- json_object
default: text
request_id:
type: string
description: >-
Unique request identifier
**Notes**:
- Passed by the client, 6-64 characters long; using UUID format is
recommended to ensure uniqueness
- If not provided, the platform will generate one automatically
minLength: 6
maxLength: 64
example: req-7f3a2c1e8b9d4f0a
user_id:
type: string
description: >-
Unique identifier of the end user
**Notes**: 6-128 characters long; using a unique identifier that
does not contain sensitive information is recommended, which can
help the platform monitor and detect abusive behavior
minLength: 6
maxLength: 128
example: user-abc123456
ChatCompletionResponse:
type: object
properties:
id:
type: string
description: Task `ID`
example: chatcmpl-a6613b56-c61c-94ba-9a9f-43d4cdc7d77a
object:
type: string
description: Response type
enum:
- chat.completion
example: chat.completion
request_id:
type: string
description: Request `ID` (returned when `request_id` is provided in the request)
example: req-7f3a2c1e8b9d4f0a
created:
type: integer
description: Request creation time, `Unix` timestamp (seconds)
example: 1777021417
model:
type: string
description: Model name
example: glm-5.3
choices:
type: array
description: The list of model responses
items:
$ref: '#/components/schemas/Choice'
usage:
$ref: '#/components/schemas/Usage'
web_search:
type: array
description: >-
Web search-related information, returned when the `web_search` tool
is used and a search is triggered
items:
$ref: '#/components/schemas/WebSearchResult'
content_filter:
type: array
description: Content safety-related information
items:
$ref: '#/components/schemas/ContentFilter'
ErrorResponse:
type: object
properties:
error:
type: object
properties:
code:
type: integer
description: HTTP status error code
message:
type: string
description: Error description
type:
type: string
description: Error type
param:
type: string
description: Related parameter name
fallback_suggestion:
type: string
description: Suggestion when error occurs
SystemMessage:
title: System Message
type: object
required:
- role
- content
properties:
role:
type: string
enum:
- system
description: Role identifier, fixed as `system`
content:
type: string
description: System prompt content, used to set the AI's role and behavior
UserMessage:
title: User Message
type: object
required:
- role
- content
properties:
role:
type: string
enum:
- user
description: Role identifier, fixed as `user`
content:
description: >-
User message content.
- **String**: plain text, supported by every model
- **Content block array**: text and images mixed, **supported only
by `glm-5.3-flash`**
Sending image content blocks to `glm-5.3` or `glm-5.2` returns an
error.
oneOf:
- type: string
title: Plain text
description: Plain-text message content
example: Hello, please introduce yourself
- type: array
title: Content block array (glm-5.3-flash only)
description: >-
Text and images mixed together. Images are passed in through
`image_url` blocks, using a public URL (recommended) or a Base64
data URL; use one `image_url` block per image.
items:
$ref: '#/components/schemas/ContentPart'
AssistantRequestMessage:
title: Assistant Message
type: object
description: Assistant message, may contain tool calls
required:
- role
properties:
role:
type: string
enum:
- assistant
description: Role identifier, fixed as `assistant`
content:
type:
- string
- 'null'
description: >-
Assistant message content
**Notes**: Used to pass historical assistant replies in multi-turn
conversations; typically `null` when `tool_calls` is present
reasoning_content:
type:
- string
- 'null'
description: >-
Historical chain-of-thought content
**Notes**: Required only when `thinking.clear_thinking=false`
(Preserved Thinking); pass back the `reasoning_content` from the
previous response as-is; by default (`clear_thinking=true`) no
passback is needed
tool_calls:
type: array
description: >-
The list of tool calls
Used to pass historical tool call information in multi-turn
conversations; when this field is provided, `content` is typically
empty
items:
type: object
required:
- id
- type
properties:
id:
type: string
description: Tool call ID
type:
type: string
enum:
- function
- web_search
description: Tool type
function:
type: object
description: Function call information, not empty when `type` is `function`
required:
- name
- arguments
properties:
name:
type: string
description: Function name
arguments:
type: string
description: Function arguments (JSON-formatted string)
ToolMessage:
title: Tool Message
type: object
required:
- role
- content
properties:
role:
type: string
enum:
- tool
description: Role identifier, fixed as `tool`
content:
type: string
description: The content of the tool call's returned result
tool_call_id:
type: string
description: >-
Indicates the tool call `ID` this message corresponds to (matching
the `id` returned in the assistant message's `tool_calls`)
FunctionTool:
title: Function Tool
type: object
required:
- type
- function
additionalProperties: false
properties:
type:
type: string
enum:
- function
default: function
description: Tool type, fixed as `function`
function:
type: object
required:
- name
- description
- parameters
properties:
name:
type: string
description: >-
The name of the function to call
**Notes**: Must consist of the characters `a-z`, `A-Z`, `0-9`,
or contain underscores and dashes; maximum length 64 characters
minLength: 1
maxLength: 64
pattern: ^[a-zA-Z0-9_-]+$
description:
type: string
description: >-
A description of the function's capability, for the model to
choose when and how to call the function
parameters:
type: object
description: >-
The function's input parameters, described as a JSON Schema
object
WebSearchTool:
title: Web Search Tool (Web Search)
type: object
required:
- type
- web_search
additionalProperties: false
properties:
type:
type: string
enum:
- web_search
default: web_search
description: Tool type, fixed as `web_search`
web_search:
type: object
required:
- enable
properties:
enable:
type: boolean
description: Whether to enable the search feature; set to `true` to enable
default: false
search_query:
type: string
description: Custom keyword that forcibly triggers a search
search_intent:
type: boolean
description: >-
Whether to perform search intent recognition; performed by
default
- `true`: Perform search intent recognition, and run the search
once a search intent is detected
- `false`: Skip intent recognition and run the search directly
count:
type: integer
description: >-
Number of results to return, in the range `1-50`, defaulting to
`10`.
The count has a direct impact on cost: the retrieved results are
merged into `prompt_tokens` and billed at the input rate, and a
value of `50` can push a single request into the tens of
thousands of input tokens. Unless you genuinely need broader
recall, keep the default.
minimum: 1
maximum: 50
default: 10
search_domain_filter:
type: string
description: >-
Domain allowlist that limits the search results (e.g.
`www.example.com`)
search_recency_filter:
type: string
description: >-
Limits the time range of the search results, defaults to
`noLimit`
enum:
- oneDay
- oneWeek
- oneMonth
- oneYear
- noLimit
default: noLimit
content_size:
type: string
description: >-
Controls the word count of web page summaries, defaults to
`medium`
- `medium`: Returns summary information, meeting basic reasoning
needs
- `high`: Maximizes context with more detailed information
enum:
- medium
- high
default: medium
result_sequence:
type: string
description: >-
The position where search results are returned (before or after
the model's reply), defaults to `after`
enum:
- before
- after
default: after
search_result:
type: boolean
description: >-
Whether to return details of the search sources in the response,
defaulting to `false`.
When set to `true`, the response carries a top-level
`web_search` array listing the sources retrieved for this
request (title, link, media source, publication date, summary,
and so on); left at the default, the field is absent from the
response.
This parameter only affects the response content — it changes
neither whether a search runs nor how it is billed.
default: false
example: true
require_search:
type: boolean
description: >-
Whether to require that the answer be returned based on search
results, defaults to `false`
default: false
search_prompt:
type: string
description: >-
The `Prompt` used to customize the processing of search results;
the default template is used when not provided
description: >-
Web search tool. Once enabled, the model can search the web when needed
and merge the results into the context.
**Billing**: the portion of the search results merged into the context
counts towards `prompt_tokens` at the input rate; the search service
itself is **billed separately per call**, settled apart from token usage
— see the pricing page. Whether a search fires is decided by the model
based on search intent, and no such charge applies when no search is
triggered.
Choice:
type: object
properties:
index:
type: integer
description: Result index
example: 0
message:
$ref: '#/components/schemas/AssistantMessage'
finish_reason:
type: string
description: >-
The reason inference terminated
- `stop`: Natural completion or a stop word was triggered
- `tool_calls`: The model triggered a function (tool call)
- `length`: Reached the token length limit
- `sensitive`: Content was blocked by safety review (please assess
and decide whether to retract public content)
- `network_error`: Model inference error
- `model_context_window_exceeded`: Exceeded the model's context
window
enum:
- stop
- tool_calls
- length
- sensitive
- network_error
- model_context_window_exceeded
example: stop
Usage:
type: object
description: Token usage statistics returned when the call ends
properties:
prompt_tokens:
type: integer
description: The number of tokens in the user input
example: 24
completion_tokens:
type: integer
description: >-
The number of output tokens (including the chain-of-thought
`reasoning_tokens` portion)
example: 346
total_tokens:
type: integer
description: Total token count = prompt_tokens + completion_tokens
example: 370
prompt_tokens_details:
type: object
description: Detailed breakdown of input tokens
properties:
cached_tokens:
type: integer
description: >-
Number of input tokens served from the context cache.
The GLM series uses an **implicit prefix cache**: repeated
requests with the same prefix hit it automatically, with no
extra parameters; the cached portion is billed at the cache
rate, markedly lower than the uncached input rate. The first
request returns 0, and subsequent requests with the same prefix
will hit.
example: 0
completion_tokens_details:
type: object
description: Detailed breakdown of output tokens
properties:
reasoning_tokens:
type: integer
description: >-
The number of tokens produced by the chain of thought (deep
thinking), counted toward `completion_tokens`
example: 321
WebSearchResult:
type: object
description: A single web search result
properties:
icon:
type: string
description: The icon of the source website
title:
type: string
description: The title of the search result
link:
type: string
description: The web page link of the search result
media:
type: string
description: The media source name of the search result web page
publish_date:
type: string
description: The website's publish time
content:
type: string
description: The cited text content of the search result web page
refer:
type: string
description: Superscript number
ContentFilter:
type: object
description: Content safety information
properties:
role:
type: string
description: |-
The stage where safety takes effect
- `assistant`: Model inference
- `user`: User input
- `history`: Historical context
enum:
- assistant
- user
- history
level:
type: integer
description: >-
Severity level `0-3`, where `0` indicates the most severe and `3`
indicates minor
minimum: 0
maximum: 3
ContentPart:
title: Content Part
type: object
description: >-
Multimodal content block. **Only `glm-5.3-flash` supports image
blocks.**
required:
- type
properties:
type:
type: string
enum:
- text
- image_url
description: |-
Content block type
- `text`: text
- `image_url`: image (`glm-5.3-flash` only)
text:
type: string
description: Text content, required when `type=text`
example: What is in this image?
image_url:
type: object
description: Image content, required when `type=image_url`
required:
- url
properties:
url:
type: string
description: >-
Image address. A public HTTPS URL (recommended) or a Base64 data
URL (`data:image/png;base64,...`) is supported
example: https://example.com/photo.jpg
AssistantMessage:
type: object
properties:
role:
type: string
description: The current conversation role, defaults to `assistant`
enum:
- assistant
example: assistant
content:
type:
- string
- 'null'
description: >-
Conversation text content
**Notes**: May be `null` when calling tools (`tool_calls`);
otherwise returns the model's reply content
example: >-
Hello! I'm GLM-5.3, and I can help you with a variety of tasks such
as conversation, reasoning, writing, and code.
reasoning_content:
type: string
description: >-
Chain-of-thought content
**Notes**: Returned when `thinking` is enabled, recording the
model's reasoning process
example: Let me first analyze this problem...
tool_calls:
type: array
description: >-
Generated tool call information (returned when the model decides to
call a tool)
items:
type: object
properties:
id:
type: string
description: Unique identifier of the tool call
type:
type: string
description: Tool call type
enum:
- function
function:
type: object
description: >-
Function call information (containing the generated function
name and JSON-formatted arguments)
properties:
name:
type: string
description: The generated function name
arguments:
type: string
description: >-
The JSON-formatted string of function call arguments;
please validate the arguments before calling the function
securitySchemes:
bearerAuth:
type: http
scheme: bearer
description: >-
##All APIs require Bearer Token authentication##
**Get API Key:**
Visit [API Key Management Page](https://evolink.ai/dashboard/keys) to
get your API Key
**Add to request header:**
```
Authorization: Bearer YOUR_API_KEY
```
````