Qwen2.5 32B Instruct
Qwen reports 83.1 on MATH and 74.5 on Arena-Hard for this 32.5B-parameter Qwen2.5 chat model. On Wiro a request holds about 17,000 input tokens, with JSON mode and function calling.
Overview
The 32B is one of two mid-size models, with a 14B, that Qwen introduced in Qwen2.5 for production use, citing strong user interest in the 10B to 30B range. The instruction-tuned 32B has 32.5B parameters, 64 layers and grouped-query attention with 40 query and 8 key-value heads. Qwen says Qwen2.5-32B beats the much larger Qwen2-72B in its evaluations.
For the chat model, Qwen's table sets it next to GPT4o-mini: 83.1 against 70.2 on MATH, 51.2 against 40.7 on LiveCodeBench, 49.5 against 40.2 on GPQA, and close on Arena-Hard (74.5 against 74.9).
On Wiro the input window is about 17,000 tokens per request, well below the card's 131,072 and roughly half the 7B's.
Choosing a variant: take the 32B when answer quality on reasoning, math or code matters more than how much text fits in one request. If prompts carry long documents, Qwen2.5 7B Instruct is the better fit. Qwen2 7B Instruct comes from the older Qwen2 line, while Qwen2.5 0.5B Instruct, Qwen2.5 3B Instruct and Qwen2.5 1.5B Instruct handle lighter jobs.
What you can build
- Review comments and patch suggestions on files or diffs that fit in about 17,000 tokens.
- Checking a worked math solution line by line and pointing to the first wrong step.
- Agent loops with several function calls per turn.
- Knowledge Q&A beyond English: Qwen reports 72.41 on TurkishMMLU and 76.55 on JMMLU.
- A grader that scores smaller models' answers against a rubric.
Inputs
A request carries a prompt or a messages list, not both. The default system prompt runs to six sentences of general safety guidance, so replace it with instructions for your task.
With a user ID and session ID, Wiro keeps the conversation between calls. About 17,000 tokens must hold everything: instructions, earlier turns, tool schemas and the new message. Anything beyond is cut, starting with old turns, then the final part of the current message. Tool schemas are never cut: if they and the system prompt leave no room for the message, the request fails with a context-length error.
Tool calling takes as many as 128 function schemas, a tool-choice mode and a parallel-calls switch. JSON mode and the usual sampling fields are there too.
Outputs
Replies are text, JSON or tool calls. The default max tokens of 0 allows 2,048 tokens for a prompt of up to 2,048 tokens and twice the prompt length for a longer one, and a higher value you set is honoured if the model's window still has room after the input. The sample covers the tower's planned 20-year life, its rescue as a radio mast and a height that shifts by about 15 cm with temperature.
Recommended settings
- Start from temperature 0.7, top-p 0.8, top-k 20 and a 1.05 repetition penalty, the values in Qwen's generation config.
- Use a separate session ID per task so unrelated history does not eat into the budget.
Limitations
- A review spanning several files often exceeds the input window, so send one file or diff per request.
- IFEval strict-prompt is 79.5, so strict multi-constraint formats still need checking.
Safety & compliance
Like the 7B, the 32B is licensed under Apache 2.0; commercial use and modified versions are permitted.
API quick start
Run Qwen2.5-32B-Instruct with a single API call.
{
"prompt": "Can you say something interesting about t…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}curl -X POST "https://api.wiro.ai/v1/Run/qwen/qwen2-5-32b-instruct" \
-H "Content-Type: application/json" \
-H "x-api-key: YOUR_WIRO_API_KEY" \
--data-binary @- <<'JSON'
{
"prompt": "Can you say something interesting about t…",
"user_id": "...",
"session_id": "...",
"system_prompt": "You are a helpful, respectful and honest …"
}
JSONOther models you might like
Explore similar llm & chat models.



