DeepSeek parameter and protocol differences

Understand DeepSeek parameter differences and usage across compatible protocols.

For DeepSeek models, adjust your chosen protocol's parameters using this page. See Chat Completions, Responses and Messages for full request and response structures.

Applicable models

Model	Applies to
deepseek-flash	Chat, Responses and Messages; image input supported
deepseek-v4-pro	Chat, Responses and Messages; no image input; Responses supports server-side search

Chat Completions differences

Field	Difference
`thinking.type`	Both models enable thinking by default and support `enabled/disabled`
`reasoning_effort`	`none` disables thinking; `low/high/max` enables it, defaulting to `high`. Compatibility levels map `minimal` to `low`, `medium/xhigh` to `high`, and `ultra` to `max`
`temperature`	Range: `0～2`; supplied values are ignored in thinking mode
`top_p`	Flash: effective thinking-mode range `0.95～1`; values below `0.95` become `0.95`. Non-thinking mode fixes it at `1`. Pro: supplied values are ignored in thinking mode, fixed at `0.95` with `tools`, otherwise `1`; effective in non-thinking mode
`max_tokens`	DeepSeek still uses this field and has not deprecated it; range: `1～393216`. Default: `8192` without thinking, `65536` with thinking, or `131072` at effort `max`
`presence_penalty`, `frequency_penalty`	Officially deprecated; supplied values have no effect
`stop`	String or string array; at most `16` stop strings
`response_format.type`	Supports `text/json_object`; JSON mode also requires explicitly requesting JSON in the prompt
`tools`, `tool_choice`	Function Tools only; `required` or a named function returns `400` in thinking mode. Disable thinking first
Historical `reasoning_content`	With thinking enabled and `tools` supplied, replay all historical assistant thinking, including turns without tool calls. Without `tools`, historical thinking is ignored
`stream_options.include_usage`	Only usable with `stream: true`. Either way, the last content chunk contains full usage, with no separate `choices: []` usage chunk. When enabled, earlier chunks have `usage` set to `null`
`user_id`	DeepSeek extension: up to `512` characters; only letters, digits, underscores and hyphens. Isolates users, caches and scheduling; exclude personal information

Responses differences

Field	Difference
`input`, `instructions`	Provide at least one; `instructions` becomes the first system message, and role `developer` is treated as `user`
`reasoning.effort`	Default: `high`; supports `none/low/high/max`, with compatibility mappings from the table above; `none` disables thinking
`temperature`, `top_p`	Values and thinking-mode behavior follow the table above; note Flash / Pro differences
`previous_response_id`, `conversation`, `store`	Stateless API; values are ignored. Replay historical input and output. Responses are fixed at `previous_response_id: null` and `store: false`
`text.format`	Supports `text/json_object/json_schema`; JSON Schema requires `name` and `schema`
`text.verbosity`, `reasoning.summary`	The former has no effect; the latter produces no summary. Historical reasoning only supports plaintext `content`, not `summary/encrypted_content`
`tools`	Supports Function Tools; Custom Tools must be named `apply_patch`, otherwise `400`. `web_search` works only on Pro and is ignored by Flash; `file_search`, `code_interpreter`, `computer_use` and `mcp` are ignored
`tool_choice`	Supports `none/auto/required` and a named Function Tool
`parallel_tool_calls`, `max_tool_calls`	Values are ignored; parallel tool calls are always enabled, and response `parallel_tool_calls` is fixed at `true`
`truncation`	Ignored; no automatic truncation. Input exceeding the context limit returns `400`
`background`, `metadata`, `include`, `prompt`	Values are ignored
`service_tier`, `safety_identifier`, `context_management`, `stream_options`	Values are ignored

Messages differences

Field	Difference
`anthropic-version`	DeepSeek ignores this Header; Anthropic version dates do not change its field rules
`thinking`, `output_config.effort`	Thinking is enabled by default; `thinking.type` supports `enabled/disabled`, while `budget_tokens` is ignored. Effort uses `low/high/max`, defaulting to `high`, with Chat's compatibility mappings; `output_config` only supports `effort`
`temperature`, `top_p`	Values and thinking-mode behavior follow the Chat table; `top_k` is ignored
`tool_choice.type`	Supports `none/auto/any/tool`; `disable_parallel_tool_use` is ignored
`metadata`	Only `user_id` is supported; other fields are ignored
`container`, `mcp_servers`, `service_tier`	Values are ignored
Text and tool results	Text-block `citations` and tool-result `is_error` are ignored; see the cache table for markers
Content blocks	Does not support `document`, `search_result`, `redacted_thinking`, `code_execution_tool_result`, `mcp_tool_use`, `mcp_tool_result` or `container_upload`

Cache differences

Field	Difference
Automatic prefix caching	Enabled by default without cache configuration; later requests can hit fully matching cached prefix units, not arbitrary matching fragments
Responses `prompt_cache_key` and `prompt_cache_retention`	Values are ignored; caching is managed automatically
Messages `cache_control`	Markers in content blocks and tool definitions are ignored
Cache usage	Chat: inspect `usage.prompt_cache_hit_tokens` and `usage.prompt_cache_miss_tokens`; Responses: inspect `usage.input_tokens_details.cached_tokens`

Media input differences

Only deepseek-flash supports images; deepseek-v4-pro does not.

Input	Protocol content block	Difference
Image	`image_url`	Chat: `image_url.url` uses a public URL or Base64 Data URL
Image	`input_image`	Responses: `image_url` uses a public URL or Base64 Data URL in user / developer messages or tool outputs. Images in system / assistant messages return `400`
Image	`image`	Messages: provide `url` for `source.type=url`; provide `media_type` and `data` for `source.type=base64`

Field	Difference
`detail`	Chat uses `image_url`; Responses uses the content block. Supports `low/high/original/auto`; `low` scales to `512×512`; the other three values are currently equivalent
Image format and size	JPEG, PNG, GIF or WebP; maximum `32 MiB` per URL / Base64 image and `48 MiB` per request body, including Base64 encoding
Image URL	Up to `8192` characters; must finish downloading within `60` seconds

Request examples

Replace MODEL_ID with the full platform model ID and use your own API Key.

curl 'https://api.agtcloud.ai/v1/chat/completions' \
  -H 'Authorization: Bearer your-api-key' \
  -H 'Content-Type: application/json' \
  --data '{
    "model": "MODEL_ID",
    "messages": [{"role": "user", "content": "Explain how to troubleshoot an API timeout in three points"}],
    "thinking": {"type": "enabled"},
    "reasoning_effort": "high",
    "max_tokens": 4096
  }'