Qwen3-Coder-Next: open weights
On 2 February 2026 Qwen (Alibaba) published the open weights of Qwen3-Coder-Next: 80 billion parameters, 3 billion active, for coding agents, under the Apache 2.0 licence.
Why it matters
A model for agentic coding with a small active part became available for local running without licence restrictions. An editorial assessment: the claims of parity with larger models are the developer's.
What is said. A hybrid architecture: 48 layers alternating Gated DeltaNet and attention, a mixture of 512 experts (10 activated and 1 shared), 262,144-token native context; non-thinking mode only. Trained on executable tasks with environments and reinforcement learning. Apache 2.0 licence per the model card. What the record does not claim. The developer's claims of over 70% on SWE-Bench Verified with the SWE-Agent scaffold and of parity with models with 10-20 times more active parameters were not independently checked. When the repository became public is not established.