Tencent Hunyuan cookbook

One problem per section. Enablement is in the tutorial. Tables are in the cheatsheet.

Streaming

Source: Hy call guide · streaming. Set stream: true. To receive usage on the last chunk, add stream_options: { include_usage: true }.

import OpenAI from 'openai'

const client = new OpenAI({
  apiKey: process.env.TOKENHUB_API_KEY,
  baseURL: 'https://tokenhub.tencentmaas.com/v1',
})

const stream = await client.chat.completions.create({
  model: 'hy3',
  messages: [{ role: 'user', content: 'Hello' }],
  stream: true,
  stream_options: { include_usage: true },
})

for await (const chunk of stream) {
  process.stdout.write(chunk.choices[0]?.delta?.content ?? '')
  if (chunk.usage) console.log('\nusage:', chunk.usage)
}

Turn thinking on

TokenHub documents two knobs on 132252:

1. thinking: { type: "enabled" } — read message.reasoning_content.
2. reasoning_effort — hy3 / hy3-preview default to low. You may set high.

Official Node sample puts thinking at the top level:

const response = await client.chat.completions.create({
  model: 'hy3',
  messages: [
    { role: 'user', content: 'A has 5 apples, gives 2 to B, then buys 3. How many remain?' },
  ],
  thinking: { type: 'enabled' },
})

const msg = response.choices[0].message
if (msg.reasoning_content) console.log('reasoning:', msg.reasoning_content)
console.log('answer:', msg.content)

Tool-call exception (official wording): on Hy3 GA, sending tools enables adaptive thinking. If you still send reasoning_effort: "low", the API maps low to high. Do not reconcile invoices against the value you typed.

Open-weight serving uses extra_body.chat_template_kwargs.reasoning_effort (no_think / low / high). That is not the same JSON as TokenHub. See the glossary.

Function calling

The call guide’s Function Calling section uses OpenAI-style tools. Keep the protocol fields; swap the function body for your backend.

const response = await client.chat.completions.create({
  model: 'hy3',
  messages: [{ role: 'user', content: 'What is the weather in Beijing today?' }],
  tools: [
    {
      type: 'function',
      function: {
        name: 'get_weather',
        description: 'Look up weather by city',
        parameters: {
          type: 'object',
          properties: {
            city: { type: 'string', description: 'City name' },
          },
          required: ['city'],
        },
      },
    },
  ],
})

Copy request/response field names from 132252. This page does not invent a second schema.

Structured output

The same call guide lists structured output as a hy3 capability. How JSON Schema is attached is defined there and in TokenHub’s API protocol page. Do not guess field names from another vendor’s response_format tutorial.

Claude Code

Official page: Call from Claude Code (Hy3). Quick start also links Cursor, OpenClaw, Cline, Kilo Code, Roo Code, and CodeBuddy Code.

Remember three facts:

1. The base URL is TokenHub’s OpenAI or Anthropic-compatible host, not api.anthropic.com.
2. The model id is hy3, not hunyuan.
3. The key is a TokenHub API key. CodeBuddy install steps belong in #78.

Do not fork a settings.json onto this site; 131903 will move.

Serve open Hy3 (only if you have the GPUs)

Source: Hy3 README. Official suggestion: 8 GPUs, H20-3e class.

vLLM (parser name hy_v3):

export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
vllm serve tencent/Hy3 \
  --tensor-parallel-size 8 \
  --speculative-config.method mtp \
  --speculative-config.num_speculative_tokens 2 \
  --tool-call-parser hy_v3 \
  --reasoning-parser hy_v3 \
  --enable-auto-tool-choice \
  --port 8000 \
  --served-model-name hy3

SGLang (parser name hunyuan):

python3 -m sglang.launch_server \
  --model tencent/Hy3 \
  --tp-size 8 \
  --tool-call-parser hunyuan \
  --reasoning-parser hunyuan \
  --speculative-num-steps 2 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 3 \
  --speculative-algorithm EAGLE \
  --port 8000 \
  --served-model-name hy3

Do not mix parser names. Recommended sampling from the README: temperature=0.9, top_p=1.0.

Recipes that do not belong here

• Choosing DeepSeek / Kimi / GLM on TokenHub.
• Yuanbao model switcher or CodeBuddy plugins.
• Treating 1729-era hunyuan-turbo / Hunyuan-T1 as the current default.