Qwen3-Coder-Next (80B-A3B)
Open 80B-A3B coding-agent model on the Qwen3-Next hybrid architecture; company says over 70% on SWE-bench Verified with 3B active parameters.
Starts from Qwen3-Next-80B-A3B-Base and trains on about 800,000 verifiable, executable coding tasks with environment feedback and RL. 262K native context; works with Claude Code, Qwen Code, Cline and others. Alibaba claims 10-20x fewer active parameters than comparable peers. Apache 2.0.
- Date
- Wednesday, 4 February 2026
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | over 70% With the SWE-Agent scaffold, per Alibaba as reported by technews.tw | company |
| Verifiable training tasks | about 800,000 Executable environments for RL | company |
Official release 2026-02-04 per press; Hugging Face weights were uploaded 2026-01-30. Technical report arXiv 2603.00729 (20 authors, 2026-02-28) confirms base and instruct weights and an agentic-training recipe; its abstract reports 'competitive' SWE-Bench and Terminal-Bench results relative to active parameters without a headline number. The 'over 70% on SWE-bench Verified' figure comes via press (using the SWE-Agent scaffold). Alibaba Cloud lists the hosted qwen3-coder-next on 2026-02-20.
Sources
- huggingface.co/Qwen/Qwen3-Coder-Next
- technews.tw/2026/02/04/qwen3-coder-next-significantly-cuts-inference-costs/
- www.alibabacloud.com/help/en/model-studio/newly-released-models
- arxiv.org/abs/2603.00729
This record was checked against its sources on 6 October 2026. How we check