Hy
Endpoint: https://mmu.vod-qcloud.com/v1认证: VOD ApiToken / Bearer
Hy(腾讯混元) 是文本生成 / 对话模型,通过兼容 OpenAI 的接口调用,支持 JSON 与 SSE 流式响应。支持 Chat Completions、Responses 与 Messages 协议,提供 hy3、hy4-preview 等版本。
基本信息
| 项目 | 取值 |
|---|---|
| 调用地址 | https://mmu.vod-qcloud.com |
| 认证 | VOD 应用签发的 ApiToken / Bearer |
| 输出 | 文本 / 结构化响应 |
| 响应方式 | JSON / SSE |
支持的协议
| 规格 | 路径 | 对象 |
|---|---|---|
| Chat Completions | POST /v1/chat/completions | 下方所有模型 |
| Responses | POST /v1/responses | 下方所有模型 |
| Messages | POST /v1/messages | 下方所有模型 |
下方模型 ID 使用对应规格。认证与公共字段遵循 API Reference。
采样选项
对本文档中列出的所有模型通用。
| 参数 | 设置 |
|---|---|
temperature | 可设置 |
top_p | 可设置 |
temperature 指定采样温度,top_p 指定累积概率范围。以一种方式为主进行调整。
各版本推理设置
| 模型 ID | 设置方式 | 停用 |
|---|---|---|
hy4-preview | thinking_enabled | 无法停用 |
hy3 | thinking_enabled | 无法停用 |
以 Chat Completions 为基准。推理内容从 reasoning_content 读取,最终答案从 content 读取。对无法停用的版本不传递停用选项。
请求参数
| 参数 | 必填 | 类型 | 说明 |
|---|---|---|---|
model | 必填 | String | 各版本选项表中的模型 ID |
messages | 必填 | Array | Chat Completions / Messages 的对话输入 |
input | 必填 | String / Array | Responses 的输入。替代 messages 使用 |
max_tokens | 可选 | Integer | Chat Completions / Messages 的生成上限。在 Messages 中为必填 |
max_output_tokens | 可选 | Integer | Responses 的生成上限 |
stream | 可选 | Boolean | true:SSE / false:JSON |
请求示例
Chat Completions
json
{
"model": "hy4-preview",
"messages": [
{
"role": "user",
"content": "Reply with exactly OK and nothing else."
}
],
"max_tokens": 512,
"stream": false
}响应
json
{
"id": "01a0de31-7e55-7dee-b6b5-2f5fb1918137",
"object": "chat.completion",
"model": "hy4-preview",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "OK",
"reasoning_content": "We need respond with exactly \"OK\" and nothing else. The user explicitly asks: \"Reply with exactly OK and nothing else.\" So final answer must be just OK, no extra characters, no punctuation? Exactly OK means two letters O and K. Ensure no newline? Usually can include newline? It says exactly OK and nothing else. Best final content: OK. Could be no trailing spaces. In chat, message body might have newline? To be safe, output just OK. No quotes. No markdown. Final: OK."
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 29,
"prompt_tokens_details": {
"audio_tokens": 0,
"cached_tokens": 0
},
"prompt_cached_tokens_details": {
"audio_tokens": 0
},
"completion_tokens": 109,
"completion_tokens_details": {
"accepted_prediction_tokens": 0,
"audio_tokens": 0,
"reasoning_tokens": 106,
"rejected_prediction_tokens": 0
},
"total_tokens": 138
}
}Responses
json
{
"model": "hy4-preview",
"input": "Reply with exactly OK and nothing else.",
"max_output_tokens": 512,
"stream": false
}响应
json
{
"id": "resp_a1c2722b42b84978bc7754f88a7d45a5",
"object": "response",
"model": "hy4-preview",
"status": "completed",
"output": [
{
"type": "reasoning",
"id": "rs_20260926225757nn1pd1f3",
"status": "completed",
"summary": [
{
"type": "summary_text",
"text": "We need respond exactly \"OK\" and nothing else. The user says: \"Reply with exactly OK and nothing else.\" So output only OK. No extra spaces? They said exactly OK and nothing else. Should we include newline? Usually just \"OK\". Could be trailing newline maybe okay? To be safe, output \"OK\" without newline? In chat, assistant message content likely \"OK\". We can just output OK. No reasoning."
}
]
},
{
"type": "message",
"id": "msg_20260926225757l9n1p3r5",
"status": "completed",
"role": "assistant",
"content": [
{
"type": "output_text",
"text": "OK"
}
]
}
],
"usage": {
"input_tokens": 29,
"output_tokens": 91,
"total_tokens": 120,
"output_tokens_details": {
"reasoning_tokens": 88
}
}
}Messages
json
{
"model": "hy4-preview",
"messages": [
{
"role": "user",
"content": "Reply with exactly OK and nothing else."
}
],
"max_tokens": 512,
"stream": false
}响应
json
{
"id": "01a0de34-91da-79f1-8dad-cdbedf2cd19f",
"model": "hy4-preview",
"type": "message",
"role": "assistant",
"content": [
{
"text": "OK",
"type": "text"
}
],
"stop_reason": "end_turn",
"usage": {
"input_tokens": 29,
"output_tokens": 111
}
}特殊设置
文本对话与代码输入传入 messages。hy3 与 hy4-preview 是不同的模型 ID。若响应中包含 reasoning_content,请与最终的 content 分开处理。