Qwen3 Coder Next

Qwen3-Coder-Next is a sparse Mixture-of-Experts model built on Qwen3-Next that carries 80 billion total parameters but activates only 3 billion during inference. Its hybrid architecture combines Gated Attention with Gated DeltaNet, allowing the model to match the coding performance of models requiring 10 to 20 times more active parameters. The design prioritizes coding agent workflows, with a 256,000-token context window, advanced long-horizon reasoning, and specialized training for navigating complex tool usage and recovering from execution failures. This combination of lightweight activation with heavy-weight capability makes it practical for agentic coding tasks in real IDE and CLI environments.

The model was trained through scaled agentic training signals, learning from large collections of verifiable coding tasks paired with executable environments that provide direct feedback. This agent-centric approach includes continued pretraining on code-focused data, supervised fine-tuning on high-quality agent trajectories, and domain-specialized expert training. Qwen3-Coder-Next achieved 70.6% on SWE-Bench and 44.3% on SWE-Bench Pro, competitive performance for its active parameter footprint. Released under an Apache 2.0 license, the open weights are available on Hugging Face in multiple variants and can run locally on consumer hardware, making elite coding agent performance accessible without expensive API costs.