Qwen3-32B
Qwen3-32B is Qwen's 32.8B dense model. It reasons step by step by default and answers directly on request, and on Wiro it can call your functions or return a JSON object.
Overview
Qwen3-32B is the 32B dense model of Qwen's Qwen3 family: 32.8B parameters in 64 layers, with 64 query heads and 8 key-value heads. One checkpoint has two modes. Thinking mode works the problem through before answering, Qwen's choice for math, code and logic. Non-thinking mode replies directly, like the older Qwen2.5 instruct models.
Wiro runs it with thinking on. Qwen documents a soft switch: /no_think in a user or system message skips the reasoning for that turn, /think brings it back, and the latest tag wins.
Choosing a variant: pick this page when your code needs a JSON object back; it is the only one of the three with JSON mode. Qwen3.5-27B is a generation newer, with 25,600 input tokens on Wiro against 17,408 here and an automatic reply budget of up to 16,384 tokens against a 4,096 cap, but it always reasons first. Qwen3.5-4B is the smallest, can switch thinking off with a field and takes 72,704 input tokens.
What you can build
- Extraction that must parse, such as turning a support email into an order object.
- Agents that call your functions and carry on once the results come back.
- Support replies and translations in the 100+ languages and dialects Qwen lists.
Inputs
Send a prompt or a messages list with the whole conversation, not both. The optional system prompt defaults to an instruction to be helpful, honest and safe. User ID and Session ID keep chat history per user and session. Sampling starts at Temperature 0.7, Top P 0.95 and Top K 0, with Do Sample on.
Function calling takes OpenAI-style definitions, a tool choice (auto, none, required or one named function) and a parallel calls switch; results go back in a follow-up request carrying the previous task's token. A Response Format of json_object turns on JSON mode.
Outputs
Text, with the reasoning and the final answer in separate segments, so the thinking can be hidden. With functions supplied, a turn can end in function calls.
Recommended settings
Qwen suggests Temperature 0.6, Top P 0.95 and Top K 20 for thinking mode, and 0.7, 0.8 and 20 for direct replies. Greedy decoding in thinking mode degrades results and can repeat endlessly, Qwen warns, so keep Do Sample on and Temperature above 0. In a messages list, drop the reasoning from earlier turns.
Limitations
- Plan for 17,408 input tokens per request. Past that, Wiro drops the oldest turns first, then cuts the end of the newest message.
- Replies stop at 4,096 tokens, reasoning included, even with a higher Max Tokens. Qwen suggests 32,768 output tokens, so on a hard prompt the thinking can use up the budget before the answer starts; /no_think helps on simple questions.
- The 32,768-token native context and the 131,072 tokens Qwen validated with YaRN are not available here.
- JSON mode fails the task when a finished reply is not a JSON object or array, after Wiro strips any code fence and repairs loose JSON such as trailing commas. A reply that runs into the 4,096-token cap comes back as it stands rather than failing, so check that it parses. Fields are not checked and there is no JSON Schema option, so validate them in your code.
Safety & compliance
Qwen releases the Qwen3-32B weights under the Apache 2.0 license.
API quick start
Run Qwen3-32B with a single API call.
{
"prompt": "Can you say something short and interesti…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/qwen/qwen3-32b" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Can you say something short and interesti…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



