Qwen2.5 Coder 32B Instruct

Overview

Qwen2.5-Coder-32B-Instruct is the instruction-tuned top end of Qwen2.5-Coder, which Qwen released in six sizes from 0.5B to 32B. The series continues training Qwen2.5 on about 5.5 trillion tokens of code-related data. This checkpoint has 32.5B parameters over 64 layers, with grouped-query attention (40 query heads, 8 key/value heads).

Qwen says the 32B model matches GPT-4o on coding. It reports 73.7 on Aider, a code-repair benchmark, and 75.2 on MdEval, a multi-language repair test where Qwen places it first among open-source models. McEval, covering more than 40 programming languages, puts it at 65.9, with Haskell and Racket called out.

Choosing a variant: this model accepts about 17,000 input tokens, while Qwen2.5 Coder 7B Instruct takes up to 32,768 and fits better when the prompt holds a large file or a long conversation. Qwen's size comparison shows scores climbing with parameter count, so pick 32B for tricky repairs and uncommon languages whenever the input fits.

What you can build

• An editor assistant that writes and fixes code inside an IDE; Qwen's announcement includes an example of the model running in Cursor.
• A bug-fix step that takes a failing snippet with its error message and returns a corrected version.
• Porting and review help for less common languages.
• Single-file HTML pages and animations from a description, like the three-body simulation in Qwen's Artifacts examples.

Inputs

Send one prompt or a messages array holding the conversation, never both. The optional system prompt holds a generic default; replace it with your coding rules. With a user_id and session_id, Wiro stores chat history so the next prompt continues the thread.

Tool use takes up to 128 OpenAI-style function definitions, with tool_choice set to auto, none, required or a named function, and parallel calls on by default. Results go back in a follow-up request that references the earlier run, and response_format can ask for a JSON object.

Sampling covers temperature (default 0.7, range 0 to 5), top_p, top_k, repetition and length penalties, minimum and maximum token counts, stop sequences and seed.

Outputs

Markdown text with code in fenced blocks, or function calls with JSON arguments when tools are attached. The sample answers the default JavaScript prompt with a one-line isPositiveNumber function (number > 0), three example calls and a note that zero returns false.

Recommended settings

• Qwen's generation config uses temperature 0.7, top_p 0.8, top_k 20 and a repetition penalty of 1.05. Wiro's defaults differ (top_p 0.95, top_k 0, penalty 1.0).
• The do_sample help text suggests turning it off for code generation, where repeatable answers matter.

Limitations

• Input over about 17,000 tokens (system prompt, history, tool definitions and new message combined) is trimmed, oldest history first, then the end of the latest message. Tool definitions are never cut, and if they and the system prompt leave no room for the message, the request fails with a context-length error.
• With max_tokens left at 0, the reply budget is 2,048 tokens for a prompt of up to 2,048 tokens and twice the prompt length for a longer one. Set max_tokens to ask for more; the only limit is the room the input leaves in the window.
• Qwen's card lists a 131,072-token context, which needs YaRN rope scaling; Wiro does not enable it.
• JSON mode has no schema enforcement, and tools must be functions.

Safety & compliance

Qwen licenses this checkpoint under Apache 2.0.