Qwen-Coder

Alibaba's Qwen team keeps its coding-specialist models in this repository, renamed from Qwen2.5-Coder when the Qwen3-Coder generation arrived. Three lines ship as open weights: a 480B mixture-of-experts flagship with 35B parameters active per token, a 30B-A3B model, and Qwen3-Coder-Next, built on the hybrid-attention Qwen3-Next 80B base with roughly 3B active parameters and trained heavily with reinforcement learning on executable tasks. All are instruct models with no separate thinking phase, a native 256K-token context that YaRN stretches toward a million, fill-in-the-middle completion, and coverage of several hundred programming languages. Tool calling relies on dedicated parsers in recent vLLM and SGLang, and the models are tuned for harnesses such as the Qwen Code terminal agent, Cline and Claude Code. Checkpoints, including FP8 and GGUF builds, sit on Hugging Face and ModelScope under Apache-2.0, so commercial use and fine-tuning are allowed. The flagship needs multi-GPU servers; the A3B variants suit a single workstation, though all weights must still fit in memory. Pick it for self-hosted coding agents.