Qwen has introduced Qwen3-Coder-Next, an open-weight language model designed for coding agents and local development. Built on Qwen3-Next-80B-A3B-Base, the model utilizes a hybrid attention and Mixture of Experts (MoE) architecture. The model's training focuses on agentic capabilities, particularly executable task synthesis, environment interaction, and reinforcement learning.

This new model emphasizes scaling agentic training signals, not just parameter size. It's trained on large, verifiable coding tasks with executable environments. The focus is on long-horizon reasoning, tool usage, and recovery from execution failures. The model demonstrates competitive performance on benchmarks such as SWE-Bench, particularly on the more challenging SWE-Bench Pro benchmark.

Qwen3-Coder-Next achieves a favorable efficiency-performance tradeoff, with a small active footprint. The model's efficiency is clear: it matches or exceeds larger open-source models in agent-centric evaluations. Demos showcase the model's capabilities in various applications, including web development and command-line interfaces.

Looking forward, Qwen aims to improve the model's reasoning, expand task support, and iterate based on user feedback. The overall goal is to enhance agent skills and improve coding agent performance. The team is focused on improving tool use and handling complex tasks. This could change the way developers approach AI-assisted coding.