Tencent's HY4 Preview mixture-of-experts model for coding agents, tool use, and long-context work.
- Modalities
- texttext
- Price
- $0.84 / $2.51/M tok
- Context
- 1M
- Providers
- 2 live
Playground
API
curl https://www.ninjachat.ai/api/v1/chat/completions \
-H "Authorization: Bearer $NINJACHAT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "hy4-preview",
"messages": [{ "role": "user", "content": "Hello!" }],
"stream": true
}'Recipes for HY4 Preview
Each request below is generated from what HY4 Preview serves on NinjaChat today — the same capabilities and parameters that GET /models reports — so it runs as written with your key.
Stream tokens
Show text as it is generated instead of waiting for the whole completion.
# export NINJACHAT_API_KEY="nj_sk_..." (Developers → Keys)
curl -N https://www.ninjachat.ai/api/v1/chat/completions \
-H "Authorization: Bearer $NINJACHAT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "hy4-preview",
"messages": [{ "role": "user", "content": "Write a haiku about deploy day." }],
"stream": true,
"stream_options": { "include_usage": true }
}'- The response is Server-Sent Events: each
data:line carries one chat.completion.chunk and the stream ends withdata: [DONE]. Both SDKs parse that for you and stop at the sentinel. - With
stream_options.include_usagethe last chunk before[DONE]has an emptychoicesarray and ausageobject with the billed token counts, so guard onchoices.lengthbefore reading a delta. - Keep
curl -Nso the buffer is not held back; the final usage chunk also carriescost_usdfor this request.
Call a tool
Let the model decide when to call your function and hand you typed arguments.
# export NINJACHAT_API_KEY="nj_sk_..." (Developers → Keys)
curl https://www.ninjachat.ai/api/v1/chat/completions \
-H "Authorization: Bearer $NINJACHAT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "hy4-preview",
"messages": [{ "role": "user", "content": "Do I need an umbrella in Lisbon today?" }],
"tools": [{
"type": "function",
"function": {
"name": "get_weather",
"description": "Current weather for a city",
"parameters": {
"type": "object",
"properties": { "city": { "type": "string" } },
"required": ["city"]
}
}
}],
"tool_choice": "auto"
}'tool_choice: "auto"lets the model answer directly when no tool is needed;"required"forces at least one call, and{ "type": "function", "function": { "name": "get_weather" } }pins a specific one.- Arguments arrive as a JSON string in
function.arguments, never as an object — parse before use. Send your result back as atoolmessage with the sametool_call_id, then call again for the natural-language answer. - Up to 32 tools per request; set
parallel_tool_calls: falsewhen your functions must run one at a time.
Get strict JSON
Have the model return an object that matches your schema, so you can parse it without cleanup.
# export NINJACHAT_API_KEY="nj_sk_..." (Developers → Keys)
curl https://www.ninjachat.ai/api/v1/chat/completions \
-H "Authorization: Bearer $NINJACHAT_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "hy4-preview",
"messages": [{ "role": "user", "content": "Extract the invoice: ACME Ltd billed 1,240.50 EUR." }],
"response_format": {
"type": "json_schema",
"json_schema": {
"name": "invoice",
"strict": true,
"schema": {
"type": "object",
"properties": {
"vendor": { "type": "string" },
"total": { "type": "number" },
"currency": { "type": "string" }
},
"required": ["vendor", "total", "currency"],
"additionalProperties": false
}
}
}
}'json_schemawithstrict: trueconstrains the output to your schema; mark every propertyrequiredand setadditionalProperties: falseso the object is exactly what you parse.{ "type": "json_object" }is the looser form — valid JSON with no schema. Whichever you use,message.contentis the JSON string; parse it, do not regex it.
Long context
Put a whole document in the prompt — the window is 1,048,576 tokens — and cap the answer.
# export NINJACHAT_API_KEY="nj_sk_..." (Developers → Keys)
DOC=$(jq -Rs . < contract.txt) # the whole file as one JSON string
curl https://www.ninjachat.ai/api/v1/chat/completions \
-H "Authorization: Bearer $NINJACHAT_API_KEY" \
-H "Content-Type: application/json" \
-d "{
\"model\": \"hy4-preview\",
\"messages\": [
{ \"role\": \"system\", \"content\": \"Answer only from the document.\" },
{ \"role\": \"user\", \"content\": $DOC },
{ \"role\": \"user\", \"content\": \"List every termination clause with its section number.\" }
],
\"max_completion_tokens\": 1200,
\"routing\": { \"caching\": \"auto\" }
}"- The context window is 1,048,576 tokens, shared between everything you send and the answer; keep the document first and the question last, and set
max_completion_tokensso a long input cannot run up an unbounded output. routing.caching: "auto"turns on provider-native prompt caching for a stable prefix. When you ask several questions over the same document, cache reads are billed at this model's cached-input rate and show up separately inusage.- Each message part is capped at 100,000 characters; split a larger document across consecutive user messages.
Pricing
- Input
- $0.84/M tokens
- Cached input
- $0.042/M tokens
- Output
- $2.51/M tokens
FAQ
How much does the HY4 Preview API cost?
HY4 Preview costs $0.84 per 1M input tokens and $2.51 per 1M output tokens on NinjaChat, billed per token with no subscription. Cached input is $0.042 per 1M tokens.
What is the context window of HY4 Preview?
HY4 Preview accepts up to 1M tokens of context per request and returns up to 64K output tokens.
Which providers serve HY4 Preview on NinjaChat?
HY4 Preview is served through Tencent Cloud TokenHub and GMI Cloud. NinjaChat routes each request to a healthy provider and fails over automatically, on one key and one balance.
What does HY4 Preview support?
HY4 Preview supports streaming, JSON mode, tool calling, reasoning, long context, and multilingual use. It is suited to multi-step reasoning and analysis, tool-calling agents and workflows, large-document and repository analysis, and structured extraction and JSON output.
How do I call HY4 Preview through the API?
Send a POST to /api/v1/chat/completions with model "hy4-preview" in the body. The endpoint is OpenAI-compatible, so the official OpenAI SDKs work after changing the base URL and key.