Qwen releases Qwen3-Coder-Next for coding agents
Alibaba's Qwen team has released Qwen3-Coder-Next, an open-weight coding model with 3B active parameters that the team says scores over 70% on SWE-Bench Verified at a fraction of typical inference cost.
The Qwen team has released Qwen3-Coder-Next, an open-weight model built for coding agents and local development. It is based on Qwen3-Next-80B-A3B-Base, a mixture-of-experts model with hybrid attention that runs about 3 billion active parameters per token.
The team says it improved the model mainly by scaling agentic training rather than model size. The model was trained on large sets of verifiable coding tasks paired with executable environments, so it could learn directly from execution feedback. The recipe has four stages: continued pretraining on code- and agent-centric data, supervised fine-tuning on agent trajectories, expert training in specific domains such as software engineering, QA and web/UX, and distillation of those experts into a single model ready for deployment. The team says the training focuses on long-horizon reasoning, tool use and recovering from failed executions.
On benchmarks, Qwen reports more than 70% on SWE-Bench Verified using the SWE-Agent scaffold. It describes results on SWE-Bench Multilingual, SWE-Bench Pro, TerminalBench 2.0 and Aider as competitive. It also says the model matches SWE-Bench Pro scores of models with 10 to 20 times more active parameters, and that its Pro results improve as more agent turns are allowed. The blog gives these comparisons mostly as charts, without other exact scores or named rival models.
Qwen shows the model running in OpenClaw, Qwen Code, Claude Code, Cline, browser-use agents and web development tools. A technical report is on GitHub. The team says the model still has "much room for improvement," and plans to work on its reasoning and decision-making next.