Qwen2.5 Coder 7B Instruct

Overview

Qwen2.5-Coder-7B-Instruct is the chat version of the 7B Qwen2.5-Coder base model. It came out in September 2024 with the 1.5B size, about two months before the rest of the series. Qwen built the series by continuing Qwen2.5 pretraining on code, and says the base models cover 92 programming languages. The checkpoint has 7.61B parameters, 28 layers, and 28 query heads that share 4 key/value heads.

For this model Qwen reports 86.7 on GSM8K, 68.7 on MMLU and 58.6 on IFEval, figures meant to show that code training kept Qwen2.5's math and general skills. Qwen also highlights its code reasoning on CRUXEval, where the model predicts what a piece of code receives and returns.

Choosing a variant: on Wiro this model reads up to 32,768 input tokens, close to double the roughly 17,000 that Qwen2.5 Coder 32B Instruct takes. Use it when a request must carry a whole module or a long chat history. When the input is short and the fix is hard, the 32B flagship is the stronger choice, since Qwen's results rise with model size.

What you can build

• Implementations written from a comment or function signature, with example calls attached.
• Walkthroughs of a complete source file for people new to a codebase.
• A debugging chat that keeps earlier turns, so each new stack trace is read against the fixes already tried.
• Small agents that pick one of your functions and fill in its arguments as JSON.

Inputs

• Prompt or messages: exactly one per request; messages holds a full conversation.
• System prompt: optional, pre-filled with a general assistant instruction you can swap for your project's conventions.
• user_id and session_id: optional labels under which Wiro saves chat history.
• Tools: up to 128 function definitions, tool_choice and a switch for several calls per turn. Results return in a follow-up request tied to the earlier run.
• response_format: plain text or a JSON object.
• Sampling: temperature 0 to 5 (default 0.7), top_p, top_k up to 100, penalties, token limits, stop sequences, seed, do_sample.

Outputs

Markdown text. With tools in play, the reply can instead name a function and give its arguments as JSON. In the sample, the isPositive function first rejects non-numbers using typeof and isNaN, then returns number > 0. Five example calls follow, including 0.1 and the string "5", which comes back false.

Recommended settings

• Qwen ships this checkpoint with temperature 0.7, top_p 0.8, top_k 20 and repetition penalty 1.1. Wiro starts top_p, top_k and the penalty at 0.95, 0 and 1.0 instead.
• Switch do_sample off for code you plan to diff or test; Wiro's help text recommends it for consistent output.

Limitations

• 32,768 input tokens per request in total, covering the system prompt, saved chat history, any tool schemas and the newest message. Past that, Wiro drops the oldest history, then cuts the tail of the newest message. Tool schemas are never trimmed, so a request whose tool schemas and system prompt leave no room for the message fails with a context-length error.
• Leaving max_tokens at 0 gives 2,048 tokens of reply when the prompt is 2,048 tokens or shorter, and twice the prompt length when it runs longer. A larger max_tokens is respected up to whatever space the input leaves in the window.
• The 131,072-token figure on Qwen's card relies on YaRN rope scaling, which is off on Wiro.
• JSON output is not checked against a schema, and only function-type tools are accepted.

Safety & compliance

Qwen publishes the 7B Instruct weights under Apache 2.0.