Tencent Hunyuan Commands & Prompt Reference
Request parameters of Tencent Hunyuan's ChatCompletions action (API version 2023-09-01), checked against TencentCloud SDK for Python 3.1.187 and the Hunyuan docs release of 2026-09-23, plus the reasoning switches of the open-weight Hy3, Hy4 preview and Hunyuan-A13B models. That release moved ChatCompletions to pre-offline status; Tencent's OpenAI-compatible interface and the image, translation and file actions are left out. Field names are PascalCase, and search stays off unless EnableEnhancement or ForceSearchEnhancement is set.
All Commands
Core request parameters (8)
| Command / Syntax | What it does | Notes |
|---|---|---|
| Model | Model name to call | Required; the reference uses hunyuan-turbos-latest as its example, and each model is billed at its own rate. |
| Messages | Conversation context | Required; up to 40 entries, oldest first; system may only lead, user and assistant alternate, and overlong context loses its oldest text. |
| Stream | Streaming response switch | Off when omitted; set true for SSE chunks in Delta that must be concatenated, which Tencent advises for lower latency. |
| StreamModeration | Streaming moderation switch | Applies only while Stream is true; on by default, it checks output in segments and may end a stream with FinishReason sensitive. |
| Temperature | Sampling temperature | Accepts 0.0 to 2.0; every model ships a recommended default, and Tencent advises leaving it unset. |
| TopP | Nucleus sampling probability threshold | Accepts 0.0 to 1.0, with larger values giving more varied wording; the model default is used when omitted. |
| Seed | Random seed for reproducible output | A non-zero integer up to 10000 makes output repeatable; Tencent warns that a poor value can hurt quality. |
| Stop | Custom stop strings that end generation | Unlike OpenAI, Hunyuan halts after the matched string rather than before it; Tencent says this may later change to match OpenAI. |
Message fields (7)
| Command / Syntax | What it does | Notes |
|---|---|---|
| Role | Message role | Takes system, user, assistant or tool. |
| Content | Message text | Plain-text body of a turn; Contents is the field for turns that mix text with images or video. |
| Contents | Multimodal content parts | Images need a Hunyuan image-to-text model and video a video-to-text model; see the product overview for which. |
| ToolCallId | Tool call reference | Set on tool-role messages to name the function call that the result answers. |
| ToolCalls | Tool calls from the model | Generated only by hunyuan-pro and hunyuan-functioncall. |
| FileIDs | Attached file IDs | One request carries at most 50 file identifiers. |
| ReasoningContent | Chain-of-thought text | Returned by Hunyuan-T1 models only; leave it out of later messages, because sending it back makes the call fail. |
Tools and reasoning parameters (4)
| Command / Syntax | What it does | Notes |
|---|---|---|
| Tools | Callable tool list | Honoured only by the hunyuan-turbos, hunyuan-t1 and hunyuan-functioncall models. |
| ToolChoice | Tool usage mode | none, auto (the default) or custom, which forces one tool; read only by hunyuan-turbo and hunyuan-functioncall. |
| CustomTool | Forced tool | Required once ToolChoice is custom; it defines the single tool the model has to call. |
| EnableThinking | Chain-of-thought reasoning switch | On when omitted; false turns reasoning off, and for now only hunyuan-a13b reads it. |
Search and enhancement switches (11)
| Command / Syntax | What it does | Notes |
|---|---|---|
| EnableEnhancement | Capability enhancement switch, such as search | Off by default since 2025-04-20 and ignored by hunyuan-lite; leaving it off lowers latency, mostly for the first streamed token. |
| ForceSearchEnhancement | Force search enhancement on | Off by default; routes each request through AI search, and the model gives a fallback reply when search returns nothing. |
| EnableSpeedSearch | Turn on the fast search variant | Off by default; when a search is triggered, the quicker variant streams its first token sooner. |
| SearchInfo | Return search results | Off by default; with it on, a response that used search includes a SearchInfo object. |
| Citation | Search citation superscript switch | Works together with EnableEnhancement and SearchInfo; searched passages get superscript marks that point into the SearchInfo list. |
| EnableMultimedia | Multimedia output switch | Allowlisted accounts only; needs EnableEnhancement on and EnableSpeedSearch off, and hunyuan-lite ignores it. |
| EnableRecommendedQuestions | Recommended follow-up questions switch | Off by default; the last chunk of the response then adds up to three RecommendedQuestions. |
| EnableDeepSearch | Deep research switch | Off by default; when a question triggers deep research, the response returns what it found. |
| EnableDeepRead | Deep document reading | Turns EnableEnhancement on by itself and handles one document in a single turn, uploaded with FilesUploads. |
| WebSearchOptions | Knowledge injection options | Object for injected knowledge, the user location, a search-status switch, and mixed image-text or music output. |
| TopicChoice | Topic passed in by the caller |
Common API parameters (5)
| Command / Syntax | What it does | Notes |
|---|---|---|
| Action | Common parameter; the API action name | Names the operation, ChatCompletions for chat; the same call can be tried in Tencent API Explorer. |
| ChatCompletions | Chat action, pre-offline | Release 53 of 2026-09-23 moved it to pre-offline status, along with the translation, thread, file and embedding actions. |
| Version | Common parameter; the API version | Fixed at 2023-09-01 for every action of the Hunyuan client. |
| Region | Common parameter; the region | Optional; the SDK adds the X-TC-Region header only when a region is configured. |
| hunyuan.tencentcloudapi.com | API endpoint host | Default host that the Hunyuan client of the Tencent Cloud SDK sends requests to. |
Open-weight model reasoning (7)
| Command / Syntax | What it does | Notes |
|---|---|---|
| reasoning_effort | Hy3 reasoning depth | Hy3 defaults to no_think, a direct answer; low and high raise the effort, high giving deep chain-of-thought for maths and code. |
| chat_template_kwargs | Chat template options | Sent inside extra_body; Hy4 preview reasons at high effort unless this object sets reasoning_effort to no_think. |
| enable_thinking=False | Fast thinking mode | Passed to apply_chat_template for Hunyuan-A13B, whose default is the slow-thinking mode with full reasoning. |
| /no_think | Skip reasoning prefix | Placed ahead of a Hunyuan-A13B prompt, it gets a direct answer for that turn without chain-of-thought. |
| /think | Force reasoning prefix | Placed ahead of a Hunyuan-A13B prompt, it makes the model reason before answering that turn. |
| <think> | Reasoning block tag | Hunyuan-A13B output opens its reasoning with this tag and closes it with the matching end tag. |
| <answer> | Final answer tag | The reply proper sits inside this pair, so the card example reads it separately from the reasoning. |
Frequently Asked Questions
Why do Hunyuan parameters look different from OpenAI's?
The hosted endpoint follows Tencent Cloud API conventions: PascalCase names, plus the common Action and Version parameters (ChatCompletions and 2023-09-01 here), which the SDK sends as X-TC-Action and X-TC-Version headers. The open-weight Hy models served with vLLM or SGLang take OpenAI-style lowercase fields instead.
How long can the conversation context be?
Messages holds at most 40 entries, ordered oldest to newest. If the combined Content exceeds the model's input limit, the earliest text is cut and the tail is kept.
Which models support tool calling and thinking?
Tools works on hunyuan-turbos, hunyuan-t1 and hunyuan-functioncall, while ToolChoice is read by hunyuan-turbo and hunyuan-functioncall. EnableThinking currently applies only to hunyuan-a13b, and ReasoningContent comes back from the Hunyuan-T1 series.
Is the ChatCompletions action still supported?
Hunyuan release 53 of 2026-09-23 moved ChatCompletions, together with the translation, glossary, thread, file and embedding actions, to pre-offline status, and the action's description now points to Tencent's OpenAI-compatible interface. The parameters on this page describe the action as last documented.
How do I switch reasoning off on the open Hy models?
Hy3 already answers directly, because reasoning_effort defaults to no_think there. Hy4 preview reasons at high effort by default, so pass no_think inside chat_template_kwargs in extra_body. Hunyuan-A13B takes enable_thinking=False in apply_chat_template or a /no_think prefix on the prompt.
How the switches combine
The search family hangs off EnableEnhancement. Citation marks are meant to be used with it and SearchInfo; EnableMultimedia only works with enhancement on and EnableSpeedSearch off; EnableDeepRead turns enhancement on by itself. ForceSearchEnhancement goes further and sends each request through AI search. Outside search, StreamModeration matters only when Stream is true, and a ToolChoice of custom needs CustomTool to define the tool.
Hosted API versus open weights
The hosted action takes PascalCase fields and a fixed Version. The SDK note added with release 53 sets 30 September 2026 as the shutdown date of the old Hunyuan console and sends callers to the OpenAI-compatible interface. The open-weight models, served with vLLM or SGLang, use the OpenAI request shape instead: Hy3 and Hy4 preview read reasoning_effort from chat_template_kwargs (no_think is the Hy3 default, high the Hy4 preview default), and both model cards recommend temperature 0.9 with top_p 1.0. Hunyuan-A13B takes enable_thinking=False or a /no_think prefix.
Sources and checks
- Tencent Cloud API, Hunyuan 2023-09-01 request models (TencentCloud SDK for Python 3.1.187, Hunyuan docs release of 2026-09-23) ChatCompletions request parameters, Message fields, WebSearchOptions
- Tencent Cloud API, Hunyuan 2023-09-01 client and action notes (TencentCloud SDK for Python 3.1.187) API version, endpoint host, ChatCompletions action description
- TencentCloud SDK for Python 3.1.187, common request client (abstract_client.py) How Action, Version and Region travel with each request
- TencentCloud SDK for Python 3.1.180 service change log (Hunyuan release 53, 2026-09-23) Pre-offline status of ChatCompletions and related actions
- Tencent Hy, Hy3 model card on Hugging Face (revision of 2026-07-20) reasoning_effort values for the open Hy3 model
- Tencent Hy, Hy4 preview model card on Hugging Face (revision of 2026-08-28) Default reasoning mode and chat_template_kwargs for Hy4 preview
- Tencent Hunyuan-A13B-Instruct model card on Hugging Face (revision of 2025-08-21) Thinking switches and output tags of Hunyuan-A13B
- Tencent Cloud API, Hunyuan 2023-09-01 API description (TencentCloud CLI 3.1.180.1, api.json) Required flags of Model and Messages
Every row on this page was matched to a line in the documentation above on . A command or parameter that could not be found there is not listed.
AI tools change their commands, parameters and model names between releases: the vendor's current documentation and the version you run have the last word. Found a wrong or missing entry? Tell us and it gets checked against the source. How we verify describes the process and its limits.
Related prompt references
DeepSeek (46) · GLM (Z.ai) (58) · Qwen (Alibaba Model Studio) (53) · All AI prompt references