GLM 5.1

zai/glm-5.1 — a llm model by Zhipu AI. Call it through the gateway’s standard OpenAI-compatible door using the catalog slug; the full parameter schema is reproduced in full below.

Catalog slug	zai/glm-5.1
Kind	llm
Vendor	Zhipu AI

How to call

Through the Sociaro SDK / OpenAI-compatible door

from sociaro_ai import Sociaro

client = Sociaro(api_key="gw_live_...")   # base_url defaults to https://api.sociaro.com

resp = client.chat.completions.create(
    model="zai/glm-5.1",          # always the catalog slug
    messages=[{"role": "user", "content": "Hello"}],
    # any parameter from the schema below is passed through unchanged
)
print(resp.choices[0].message.content)

TypeScript: client.chat.completions.create({ model: "zai/glm-5.1", messages: [...] }) — see the TypeScript SDK.

Parameters — full schema

The complete, one-to-one API schema for this model, transcribed from the official source. Nothing omitted. Examples in the schema may use a representative model of the family — for this model use the slug zai/glm-5.1 as shown in How to call above.

This is the gateway’s standard OpenAI-compatible Chat Completions reference. The models documented here (ByteDance Seed, DeepSeek, OpenAI gpt-oss, Zhipu GLM) are all reached through the same chat door using their catalog slug as model. The parameters below follow the OpenAI Chat Completions schema; the per-model capability notes describe behaviour of the underlying models, not any particular routing provider.

---

Endpoint

POST https://api.sociaro.com/v1/chat/completions
Content-Type: application/json
Authorization: Bearer $SOCIARO_API_KEY

• Auth: Bearer API key (Authorization: Bearer $SOCIARO_API_KEY).
• model is the catalog slug for the model you want (e.g. bytedance/seed-2-0-lite, deepseek/deepseek-v4-pro, openai/gpt-oss-120b, zai/glm-4-7, zai/glm-5.1).
• Largely OpenAI-compatible. Extras beyond vanilla OpenAI Chat: thinking{type}, reasoning_effort, service_tier, max_completion_tokens, messages.reasoning_content, messages.encrypted_content, image_url.image_pixel_limit, image_url.detail=xhigh, video_url.fps, audio input_audio part, file part, file file_id inputs, choices.moderation_hit_type, usage.prompt_tokens_details.audio_tokens/audio_cached_tokens.
• This is the union superset across the Seed / DeepSeek / GPT-OSS / GLM families served via this door. The gateway treats bodies transparently, so per-model carve-outs are documentation notes, not request validation.

---

Per-model coverage map (capability matrix)

thinking.type support and default (per-model deep-reasoning capability):

Catalog slug	thinking.type supported	Default	Notes
deepseek/deepseek-v4-pro	enabled, disabled	enabled	reasoning_effort incl. max
deepseek/deepseek-v4-flash	enabled, disabled	enabled	reasoning_effort incl. max
deepseek/deepseek-v3-2	enabled, disabled	disabled	the only listed model defaulting to disabled; supports messages.reasoning_content
bytedance/seed-2-0-pro	enabled, disabled	enabled	reasoning_effort; thinking-summary model
bytedance/seed-2-0-lite	enabled, disabled	enabled	reasoning_effort; thinking-summary; encrypted_content supported
bytedance/seed-2-0-mini	enabled, disabled	enabled	reasoning_effort; thinking-summary; file_id input supported
bytedance/seed-1-8	enabled, disabled	enabled	reasoning_effort; keeps reasoning_content in history
bytedance/seed-1-6	enabled, disabled	enabled	seed-1.6+: honors parallel_tool_calls=false & tool_choice
bytedance/seed-1-6-flash	enabled, disabled	enabled
openai/gpt-oss-120b	enabled, disabled, auto	enabled	only model listing auto; OpenAI-style reasoning_effort
zai/glm-4-7	enabled, disabled	enabled

reasoning_effort supported models (chain-of-thought length control): deepseek/deepseek-v4-pro, deepseek/deepseek-v4-flash, bytedance/seed-2-0-lite, bytedance/seed-2-0-mini, bytedance/seed-2-0-pro, bytedance/seed-1-8. Value max works ONLY on deepseek/deepseek-v4-pro & deepseek/deepseek-v4-flash (ignored if passed elsewhere).

Thinking-summary models (return reasoning_content summary + encrypted_content instead of raw CoT; enabled by default): bytedance/seed-2-0-pro, bytedance/seed-2-0-mini, bytedance/seed-2-0-lite.

Which models are deep-reasoning-capable (and hence drop logprobs/top_logprobs/logit_bias/stop) varies by model. Always confirm a specific model’s capability before relying on those four params.

Critical per-model differences (request-side carve-outs)

Param / field	Carve-out
thinking.type=auto	A valid value globally, but only openai/gpt-oss-120b lists auto support. Non-thinking models ignore thinking.
thinking default	{"type":"enabled"} globally; but deepseek/deepseek-v3-2 defaults to disabled. Support & default vary by model.
reasoning_effort	Only the models listed above accept it; max only on deepseek/deepseek-v4-pro & deepseek/deepseek-v4-flash.
frequency_penalty, presence_penalty	NOT supported by seed-1.8 and seed-2.0 series (ignored there). Supported elsewhere, range [-2,2].
logprobs, top_logprobs, logit_bias	NOT supported by deep-reasoning-capable models.
stop	NOT supported by deep-reasoning models; max 4 strings otherwise.
parallel_tool_calls=false	Only honored by seed-1.6 and later.
tool_choice	Only seed-1.6 and later support this parameter.
messages.reasoning_content (assistant input)	Only seed-1.8, seed-2.0, deepseek-v3.2. Multi-turn: models before 251228 should have it stripped from history; seed-1.8+ keep it.
messages.encrypted_content (assistant input)	Supported on bytedance/seed-2-0-lite, bytedance/seed-2-0-mini, bytedance/seed-2-0-pro. Takes precedence over reasoning_content; invalid/tampered value → service returns “Invalid signature”.
image_url.image_pixel_limit.max_pixels	≤ 4 014 080 for models before seed-1.8; ≤ 9 031 680 for seed-1.8 & seed-2.0.
image_url.image_pixel_limit.min_pixels	≥ 3136 (pre-1.8); ≥ 1764 (seed-1.8 & seed-2.0).
image_url.detail	Valid values low / high / xhigh; overridden by image_pixel_limit when both set.
*.file_id inputs (image/video/audio/file via Files API or TOS bucket)	Supported on: bytedance/seed-2-0-mini, bytedance/seed-2-0-lite, bytedance/seed-2-0-pro.
choices.moderation_hit_type (response)	Returned only by visual-understanding models with ModerationStrategy=Basic set on the endpoint.
service_tier (response value)	scale when a TPM guarantee package was used, default when not. (Request side uses auto/default — different value set.)
response_format json_object / json_schema	Beta; only specific models (structured output).

---

Request body — top-level shape

{
  "model": "string (required)",
  "messages": [ /* required, array of message objects */ ],
  "thinking": { "type": "enabled | disabled | auto" },
  "stream": false,
  "stream_options": { "include_usage": false },
  "max_tokens": 4096,
  "max_completion_tokens": 0,
  "service_tier": "auto | default",
  "stop": "string | [\"s1\",\"s2\"] | null",
  "reasoning_effort": "minimal | low | medium | high | max",
  "response_format": { "type": "text | json_object | json_schema", "json_schema": { } },
  "frequency_penalty": 0,
  "presence_penalty": 0,
  "temperature": 1,
  "top_p": 0.7,
  "logprobs": false,
  "top_logprobs": 0,
  "logit_bias": { "<token_id>": 0 },
  "tools": [ { "type": "function", "function": { } } ],
  "parallel_tool_calls": true,
  "tool_choice": "none | auto | required | {object}"
}

---

Request parameters — top-level

Name	Type	Required	Default	Allowed / Range	Description
model	string	required	—	zai/glm-5.1	Model to call — the catalog slug for this model.
messages	object[]	required	—	see message types below	List of chat messages. Supported content modalities (text/image/video/audio/file) vary by model.
thinking	object	optional	{"type":"enabled"}	see thinking.type	Enables/disables deep-thinking mode. Support & default vary by model; deepseek/deepseek-v3-2 defaults to disabled.
thinking.type	string	required (within object)	enabled	enabled, disabled, auto	enabled=always think first; disabled=answer directly (no CoT); auto=model decides, skipping CoT for simple questions (only openai/gpt-oss-120b lists auto).
stream	boolean / null	optional	false	true, false	false=full response at once; true=SSE blocks per SSE protocol ending with data: [DONE].
stream_options	object / null	optional	null	—	Streaming options; only meaningful when stream=true.
stream_options.include_usage	boolean / null	optional	false	true, false	true=extra block before data:[DONE] whose usage is the whole-request token usage and whose choices is an empty array.
max_tokens	integer / null	optional	4096	varies by model	Max length of the model response in tokens. Response = output − chain-of-thought (if any). Also bounded by context length.
max_completion_tokens	integer / null	optional	—	[1, 65536]	Max output length including response + chain-of-thought. When set, the model ignores the max_tokens default and generates up to this value. Cannot be set together with max_tokens. Preferred for deep-reasoning models.
service_tier	string / null	optional	auto	auto, default	auto=use TPM guarantee package if the endpoint has quota (faster / higher availability), else default tier; default=never use TPM package even if quota exists.
stop	string / string[] / null	optional	null	up to 4 strings	Generation stops when a stop string is encountered; the stop string is excluded from the output. Not supported by deep-reasoning models.
reasoning_effort	string / null	optional	medium	minimal, low, medium, high, max	CoT depth. minimal=no reasoning/answer directly; low=light/fast; medium=balanced; high=deep analysis; max=only deepseek/deepseek-v4-pro & deepseek/deepseek-v4-flash (ignored elsewhere). Interacts with thinking.type; only applies to original CoT, not the thinking summary.
response_format	object	optional (beta)	{"type":"text"}	see sub-table	Output format. json_object/json_schema are beta, model-dependent.
frequency_penalty	float / null	optional	0	[-2.0, 2.0]	Positive values penalize tokens by frequency, reducing repetition. Not supported by seed-1.8 / seed-2.0 series.
presence_penalty	float / null	optional	0	[-2.0, 2.0]	Positive values penalize tokens by presence, encouraging new topics. Not supported by seed-1.8 / seed-2.0 series.
temperature	float / null	optional	1	[0, 2]	Sampling temperature. 0=only top-logprob token; higher=more random, lower=more deterministic. Adjust either this or top_p.
top_p	float / null	optional	0.7	[0, 1]	Nucleus-sampling threshold (tokens whose cumulative prob ≥ top_p). 0=only top token; 0.1=top 10% probability mass. Adjust either this or temperature.
logprobs	boolean / null	optional	false	true, false	Return logprobs of tokens. Not supported by deep-reasoning-capable models.
top_logprobs	integer / null	optional	0	[0, 20]	Number of most-likely tokens (with logprob) per position. Requires logprobs=true. Not supported by deep-reasoning models.
logit_bias	map / null	optional	null	values [-100, 100]	Map of token ID (from the Tokenization API vocabulary) → bias. -100=never use; 100=only use; -1/1=discourage/encourage. Actual effect varies by model. Not supported by deep-reasoning models.
tools	object[] / null	optional	null	see sub-table	Tools the model may call (function calling). Model-dependent.
parallel_tool_calls	boolean	optional	true	true, false	true=allow multiple tools in one response; false=≤1 tool. false only honored by seed-1.6+.
tool_choice	string / object	optional	none if no tools, auto if tools present	none, auto, required, or {type,function}	Controls tool usage. Only seed-1.6+ support this param. none=no tool; required=must call a tool; auto=model decides; object=force a specific function.

response_format variants

Name	Type	Required	Default	Allowed	Description
response_format.type	string	required	text	text / json_object / json_schema	Output format selector. json_object/json_schema beta + model-gated.
response_format.json_schema	object	required when type=json_schema	—	—	JSON-schema definition wrapper.
response_format.json_schema.name	string	required	—	—	Name of the user-defined schema.
response_format.json_schema.description	string / null	optional	—	—	Purpose description; guides how the model formats the response.
response_format.json_schema.schema	object	required	—	JSON Schema object	The response structure definition (JSON Schema).
response_format.json_schema.strict	boolean / null	optional	false	true/false	true=strictly follow schema; false=best-effort.

tools[] (function calling)

Name	Type	Required	Allowed	Description
tools.type	string	required	function	Tool type (only function).
tools.function	object	required	—	Function definition.
tools.function.name	string	required	—	Function name (case-sensitive).
tools.function.description	string	optional	—	Helps the model decide whether to call it.
tools.function.parameters	object	optional	JSON Schema	Function args as a JSON Schema object (type/properties/required). All names case-sensitive; must be a valid JSON Schema object.

tool_choice object form

Name	Type	Required	Allowed	Description
tool_choice.type	string	required	function	Tool type.
tool_choice.function	object	required	—	Target tool.
tool_choice.function.name	string	required	—	Name of the tool to force.

---

messages[] — message objects

Four roles: system, user, assistant (model message), tool.

System message

Name	Type	Required	Allowed	Description
messages.role	string	required	system	Role.
messages.content	string / object[]	required	plaintext or multimodal parts	Instructions/context (role, context, etc.).

User message

Name	Type	Required	Allowed	Description
messages.role	string	required	user	Role.
messages.content	string / object[]	required	plaintext or multimodal parts	User content; supported modalities vary by model.

Assistant (model) message — for multi-turn history / prefill

At least one of content or tool_calls is required.

Name	Type	Required	Allowed	Description
messages.role	string	required	assistant	Role.
messages.content	string / array	conditional	—	Prior model content.
messages.reasoning_content	string	optional	—	Chain-of-thought content. Only seed-1.8, seed-2.0, deepseek-v3.2. Older seed-1.6 snapshots: strip from history; seed-1.8+: keep.
messages.encrypted_content	string	optional	—	Encrypted/compressed reasoning. Supported on bytedance/seed-2-0-lite, bytedance/seed-2-0-mini, bytedance/seed-2-0-pro. Takes precedence over reasoning_content; tampered/invalid → “Invalid signature”.
messages.tool_calls	object[]	conditional	—	Prior tool-call messages emitted by the model.
messages.tool_calls.id	string	required	—	Tool-call ID (model-generated).
messages.tool_calls.type	string	required	function	Tool type.
messages.tool_calls.function	object	required	—	Function info.
messages.tool_calls.function.name	string	required	—	Function name.
messages.tool_calls.function.arguments	string	required	JSON string	Generated args (may be invalid — validate before use).

Tool message — function-calling result

Name	Type	Required	Allowed	Description
messages.role	string	required	tool	Role.
messages.content	string / array	required	—	Tool output.
messages.tool_call_id	string	required	—	The model-generated tool-call ID this result answers (associates the tool result with the originating call).

---

Multimodal content[] parts (within user/system content arrays)

Text part

Name	Type	Required	Allowed	Description
messages.content.type	string	required	text	Part type.
messages.content.text	string	required	—	Text content.

Image part

Pass exactly one of file_id / url.

Name	Type	Required	Default	Allowed / Range	Description
messages.content.type	string	required	—	image_url	Part type.
messages.content.image_url	object	required	—	—	Image content.
messages.content.image_url.file_id	string	conditional	—	—	Files-API file ID. File type must match type; status must be active; project must match the API key’s project. TOS-bucket file_id supported on bytedance/seed-2-0-mini, bytedance/seed-2-0-lite, bytedance/seed-2-0-pro.
messages.content.image_url.url	string	conditional	—	image link or base64	Image URL or base64 string.
messages.content.image_url.detail	string / null	optional	model-default	low, high, xhigh	Image granularity; overridden by image_pixel_limit when both set.
messages.content.image_url.image_pixel_limit	object / null	optional	null	total pixels must be within [196, 36000000]	Pixel bounds; image outside the range is proportionally resized. Takes precedence over detail. If a sub-field is unset, falls back to the value implied by detail.
messages.content.image_url.image_pixel_limit.max_pixels	integer	optional	from detail	≤ 4 014 080 (pre-1.8); ≤ 9 031 680 (seed-1.8 & 2.0); range (min_pixels, ceiling]	Downscale ceiling: images with more pixels scaled down below it.
messages.content.image_url.image_pixel_limit.min_pixels	integer	optional	from detail	≥ 3136 (pre-1.8); ≥ 1764 (seed-1.8 & 2.0); range [floor, max_pixels)	Upscale floor: images with fewer pixels scaled up above it.

Video part

Pass one of file_id / url. Audio comprehension of video content is not supported in user-message video parts.

Name	Type	Required	Default	Range	Description
messages.content.type	string	required	—	video_url	Part type.
messages.content.video_url	object	required	—	—	Video content.
messages.content.video_url.file_id	string	conditional	—	—	Files-API file ID (same constraints as image file_id; same TOS/model gating).
messages.content.video_url.url	string	conditional	—	video link or base64	Video source.
messages.content.video_url.fps	float / null	optional	1	[0.2, 5]	Frames extracted per second. Higher=more detail/more tokens; lower=fewer tokens/faster.

Audio part

type = input_audio. Pass one of file_id / url / data.

Name	Type	Required	Allowed	Description
messages.content.type	string	required	input_audio	Part type.
messages.content.input_audio	object	required	—	Audio content.
messages.content.input_audio.file_id	string	conditional	—	Files-API file ID (same constraints / model gating).
messages.content.input_audio.url	string	conditional	—	Audio URL.
messages.content.input_audio.data	string	conditional	base64	Base64-encoded audio.
messages.content.input_audio.format	string	required when data used	see below	MIME format of the audio.

format MIME values:

• Audio-only: .mp3=audio/mpeg, .wav=audio/wav, .aac=audio/aac, .m4a=audio/mp4
• Embedded-in-video also: .mp3=audio/mpeg, .wav=audio/wav, .aac=audio/aac, .m4a=audio/mp4, .pcm=audio/L16, .ac3=audio/ac3, .alac=audio/mp4

Audio limits: file size ≤ 25 MB; total audio-only duration ≤ 120 min per request (embedded-in-video audio duration is not counted).

File part (PDF only)

type = file. Pass one of file_id / file_data / file_url. PDF only.

Name	Type	Required	Allowed	Description
messages.content.type	string	required	file	Part type.
messages.content.file	object	required	—	File content (PDF only).
messages.content.file.file_id	string	conditional	—	Files-API file ID (same constraints / model gating).
messages.content.file.file_data	string	conditional	base64	Base64 file content; single file ≤ 50 MB.
messages.content.file.filename	string	required when file_data used	—	Filename.
messages.content.file.file_url	string	conditional	—	Accessible URL; file ≤ 50 MB.

---

Response — non-streaming (object: chat.completion)

{
  "id": "string",
  "model": "string",
  "service_tier": "scale | default",
  "created": 1749110144,
  "object": "chat.completion",
  "choices": [
    {
      "index": 0,
      "finish_reason": "stop | length | content_filter | tool_calls",
      "logprobs": null,
      "message": {
        "role": "assistant",
        "content": "string",
        "reasoning_content": "string | null",
        "tool_calls": [
          { "id": "string", "type": "function",
            "function": { "name": "string", "arguments": "json-string" } }
        ]
      },
      "moderation_hit_type": "severe_violation | violence | null"
    }
  ],
  "usage": {
    "total_tokens": 55,
    "prompt_tokens": 20,
    "prompt_tokens_details": { "cached_tokens": 0, "audio_tokens": 0, "audio_cached_tokens": 0 },
    "completion_tokens": 35,
    "completion_tokens_details": { "reasoning_tokens": 0 }
  }
}

Response fields (non-streaming)

Name	Type	Description
id	string	Unique request identifier.
model	string	Model name + version used.
service_tier	string	scale=TPM guarantee package quota used; default=not used.
created	integer	Unix timestamp (seconds) of request creation.
object	string	chat.completion.
choices	object[]	Model outputs.
choices.index	integer	Index in choices.
choices.finish_reason	string	stop=ended or hit a stop string; length=hit max_tokens / max_completion_tokens / context_window; content_filter=intercepted by moderation; tool_calls=a tool was called.
choices.message.role	string	assistant.
choices.message.content	string	Generated message.
choices.message.reasoning_content	string / null	Chain-of-thought during problem-solving (deep-reasoning models only).
choices.message.tool_calls	object[] / null	Tool calls generated by the model.
choices.message.tool_calls.id	string	Tool-call ID.
choices.message.tool_calls.type	string	function.
choices.message.tool_calls.function.name	string	Function name.
choices.message.tool_calls.function.arguments	string	JSON-string args (may be invalid — validate before use).
choices.logprobs	object / null	Logprob detail (null if logprobs not requested).
choices.logprobs.content	object[] / null	Per-token logprob for each content element.
choices.logprobs.content[].token	string	Token text.
choices.logprobs.content[].bytes	integer[] / null	UTF-8 byte values ([] if the token has none).
choices.logprobs.content[].logprob	float	Token logprob.
choices.logprobs.content[].top_logprobs[]	object[]	Most-probable tokens at the position (each: token / bytes / logprob). May return fewer than requested.
choices.moderation_hit_type	string / null	severe_violation / violence. Visual-understanding models only, with ModerationStrategy=Basic on the endpoint.
usage.total_tokens	integer	Input + output tokens.
usage.prompt_tokens	integer	Input tokens.
usage.prompt_tokens_details.cached_tokens	integer	Cached input tokens (all modalities — text, audio, etc.).
usage.prompt_tokens_details.audio_tokens	integer	Audio-input tokens.
usage.prompt_tokens_details.audio_cached_tokens	integer	Cached audio-input tokens.
usage.completion_tokens	integer	Output tokens.
usage.completion_tokens_details.reasoning_tokens	integer	Chain-of-thought tokens. (For thinking-summary models, billing still counts the original CoT tokens.)

---

Response — streaming (object: chat.completion.chunk)

Same top-level fields (id, model, service_tier, created) but:

• object = chat.completion.chunk
• each chunk uses choices[].delta (instead of message)
• stream ends with data: [DONE]
• usage is null by default; set stream_options.include_usage=true to get a final usage chunk (with empty choices[])

Name	Type	Description
choices.index	integer	Index in choices.
choices.finish_reason	string	Same values as non-streaming.
choices.delta.role	string	assistant.
choices.delta.content	string	Incremental content.
choices.delta.reasoning_content	string / null	Incremental reasoning (deep-reasoning models). On thinking-summary models (bytedance/seed-2-0-lite, bytedance/seed-2-0-mini, bytedance/seed-2-0-pro) this returns a summary of the reasoning, not the raw CoT.
choices.delta.encrypted_content	string	Full encrypted/compressed reasoning emitted in one chunk after CoT streaming finishes and before the response begins; in that chunk both content and reasoning_content are empty. Supported on bytedance/seed-2-0-lite, bytedance/seed-2-0-mini, bytedance/seed-2-0-pro.
choices.delta.tool_calls[]	object[] / null	Incremental tool calls (id / type=function / function.name / function.arguments).
choices.logprobs	object / null	Same structure as non-streaming.
choices.moderation_hit_type	string / null	Same as non-streaming.
usage	object / null	Same structure as non-streaming; only present when include_usage=true. In streaming, usage.prompt_tokens_details.cached_tokens is documented as 0.

---

Minimal example

curl https://api.sociaro.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SOCIARO_API_KEY" \
  -d '{
    "model": "bytedance/seed-2-0-lite",
    "messages": [
      { "role": "system", "content": "You are a helpful assistant." },
      { "role": "user", "content": "Hello!" }
    ]
  }'

Minimal response (verbatim shape from the doc):

{
  "choices": [{
    "finish_reason": "stop",
    "index": 0,
    "logprobs": null,
    "message": { "content": "Hello! How are you? ...", "role": "assistant" }
  }],
  "created": 1749110144,
  "id": "0217491101427332b0b8a623b95f3e6209666650b67ac353d7ad2",
  "model": "bytedance/seed-2-0-lite",
  "service_tier": "default",
  "object": "chat.completion",
  "usage": {
    "completion_tokens": 35,
    "prompt_tokens": 20,
    "total_tokens": 55,
    "prompt_tokens_details": { "cached_tokens": 0 },
    "completion_tokens_details": { "reasoning_tokens": 0 }
  }
}

Full example (thinking + tools + streaming usage + multimodal + sampling controls)

curl https://api.sociaro.com/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer $SOCIARO_API_KEY" \
  -d '{
    "model": "deepseek/deepseek-v4-pro",
    "messages": [
      { "role": "system", "content": "You are a helpful assistant." },
      { "role": "user", "content": [
          { "type": "text", "text": "Describe this image and call get_weather." },
          { "type": "image_url",
            "image_url": { "url": "https://example.com/a.jpg", "detail": "high",
                           "image_pixel_limit": { "max_pixels": 9031680, "min_pixels": 1764 } } }
        ]
      }
    ],
    "thinking": { "type": "enabled" },
    "reasoning_effort": "max",
    "max_completion_tokens": 8192,
    "service_tier": "auto",
    "temperature": 0.7,
    "top_p": 0.9,
    "stop": ["\n\nUser:"],
    "logprobs": false,
    "response_format": { "type": "text" },
    "tools": [
      { "type": "function",
        "function": {
          "name": "get_weather",
          "description": "Get current weather for a city.",
          "parameters": {
            "type": "object",
            "properties": { "city": { "type": "string", "description": "City name" } },
            "required": ["city"]
          }
        }
      }
    ],
    "parallel_tool_calls": false,
    "tool_choice": "auto",
    "stream": true,
    "stream_options": { "include_usage": true }
  }'

Full-example notes: max_completion_tokens is used instead of max_tokens (mutually exclusive). reasoning_effort:"max" only takes effect on deepseek/deepseek-v4-pro / deepseek/deepseek-v4-flash. parallel_tool_calls:false and tool_choice require seed-1.6+ (shown on deepseek to illustrate field placement; deepseek honoring depends on the model’s tool support). stop is unsupported on deep-reasoning models and is ignored there; logprobs/top_logprobs/logit_bias likewise. On deepseek/deepseek-v4-pro thinking defaults to enabled; on deepseek/deepseek-v3-2 it would default to disabled.

Audio / file input snippets

Audio (base64) part:

{ "type": "input_audio",
  "input_audio": { "data": "<base64>", "format": "audio/mpeg" } }

PDF file part:

{ "type": "file",
  "file": { "file_data": "<base64>", "filename": "doc.pdf" } }

Files-API file_id part (bytedance/seed-2-0-mini / bytedance/seed-2-0-lite / bytedance/seed-2-0-pro only):

{ "type": "image_url", "image_url": { "file_id": "<files-api-id>" } }