Qwen2.5-Coder¶

Development line: project:qwen2-5-coder · thread qwen2-5-coder-launch
Last event: 2024-11-12 · 2 dated since 2024-11-12 · Researched: 2026-09-04 · confidence: high

What it is¶

Qwen2.5-Coder is a family of six base and instruction-tuned code LLMs from 0.5B to 32B for completion, chat, code repair, and fine-tuning.

• Generation: generates code.
• Reasoning: reasons through programming tasks.
• Repair: fixes bugs in code.
• Fill-in-the-middle: completes code between prefix and suffix.

Context reaches up to 128K, but the standard configuration sets 32K. The family fits local and self-hosted tasks, but Qwen directs new agentic development to Qwen3-Coder.

Development line¶

• 2024-11-12 — Qwen released the Qwen2.5-Coder model family. To the previously available 1.5B and 7B, Qwen added 0.5B, 3B, 14B, and 32B in base and instruct variants.
• 2024-11-12 — Qwen made a Qwen2.5-Coder Artifacts experience available. To the previously available 1.5B and 7B, Qwen added 0.5B, 3B, 14B, and 32B in base and instruct variants.

What changed¶

2024-11-12 — Qwen opened the full Qwen2.5-Coder lineup: to the previously available 1.5B and 7B, Qwen added 0.5B, 3B, 14B, and 32B in base and instruct variants. 2025-07-22 — Qwen released Qwen3-Coder as the next generation for agentic coding; the original Qwen2.5-Coder repository now redirects to it.

How to use this¶

From 2024-11-12, practitioners could evaluate Qwen2.5-Coder through its released model collection, demo, repository, and Artifacts experience rather than relying only on a general chat interface.

1. For interactive tasks, choose an instruction variant of suitable size, load the tokenizer and model through Transformers, and build the request with apply_chat_template. — https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct
2. For a self-hosted API, serve the chosen checkpoint through vLLM; the server exposes an OpenAI-compatible /v1/chat/completions endpoint. — https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct
3. For completion or FIM, run the base model; for normal dialogue, use Instruct rather than base. — https://huggingface.co/Qwen/Qwen2.5-Coder-7B

Best practices¶

• Do not use base models for conversation without custom post-training; choose Instruct for an assistant. — https://huggingface.co/Qwen/Qwen2.5-Coder-7B
• Enable YaRN only when context above 32K is truly needed: static scaling can hurt short requests. — https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct
• Update Transformers to at least version 4.37, or loading the Qwen2 architecture fails with KeyError. — https://huggingface.co/Qwen/Qwen2.5-Coder-7B

Superseded by this¶

• 2025-07-22 — Qwen3-Coder became Qwen's current line for agentic coding; Qwen2.5-Coder remains an available set of checkpoints, but its original GitHub repository redirects to Qwen3-Coder.

Still unknown¶

• The current status of the public Qwen2.5-Coder-Artifacts Space is paused; its availability history and fitness as a working product remain unconfirmed.
• The 3B license differs: it uses the Qwen Research License, whereas 0.5B, 1.5B, 7B, 14B, and 32B use Apache 2.0.

Sources¶

Agent brief¶

• Subject: project:qwen2-5-coder, thread qwen2-5-coder-launch, 2 dated events 2024-11-12 → 2024-11-12.
• Practical note: From 2024-11-12, practitioners could evaluate Qwen2.5-Coder through its released model collection, demo, repository, and Artifacts experience rather than relying only on a general chat interface.
• Confidence: high. Dated supersedes above are the authority for what is obsolete.