Qwen2.5-Coder

Qwen2.5-Coder is a family of coding language models for developers who want a local coding assistant or a model for a self-hosted code agent. The 32B instruction-tuned model supports conversational requests about code.

• Code generation and fixes: Generate code, reason through its behavior and ask for corrections to errors.
• Explanations beyond code: Use the model's mathematics and general language capabilities for related questions within a coding conversation.
• Model size choices: The family includes smaller models alongside the 32B model, so developers can choose a size for their deployment.
• Long code inputs: The 32B instruction model supports up to 128K tokens with YaRN context extension in compatible serving frameworks; its default context is about 32K tokens.

The 32B instruction model is open source under Apache 2.0, with weights in Safetensors format. It works with Hugging Face Transformers for local execution and supports vLLM for serving on your own infrastructure. Static YaRN scaling in vLLM can affect performance on shorter inputs, so the larger context window has a trade-off.

Hosted inference through Nscale is also available on Hugging Face. That option runs on the provider's infrastructure rather than your own hardware.