Top AI Product

Alibaba’s Qwen team shipped Qwen3-Coder-Next, an open-weight model built for coding agents rather than chat. It’s ultra-sparse — built on the Qwen3-Next-80B-A3B base with hybrid attention and MoE, activating only a fraction of its parameters per token for roughly 10x higher throughput on repo-scale tasks.

## The benchmarks

It clears 70% on SWE-Bench Verified using the SWE-Agent scaffold, and stays competitive on the harder SWE-Bench Pro and Multilingual splits, plus TerminalBench 2.0 and Aider. For an open model you can self-host, landing in that range on real bug-fixing tasks is the number that matters to anyone building a coding agent they would rather own than rent.

## How it was trained

The interesting part is the data pipeline. Qwen mined 800,000 verifiable coding tasks — real bug-fixing scenarios pulled from GitHub pull requests, each paired with a fully executable environment. Training ran on MegaFlow, a Kubernetes-based orchestration system where every agentic task is a three-stage loop: agent rollout, evaluation, post-processing. It’s reinforcement learning on coding at scale, and unlike the closed frontier coders it competes with, the weights are open.