Qwen3.6-35B-A3B and Qwen3.6-27B
Open 35B-A3B MoE and 27B dense models tuned for agentic coding; SWE-bench Verified 73.4 and 77.2, with 'preserve thinking' across turns.
Better frontend and repository-level coding, plus a 'preserve_thinking' option that keeps earlier reasoning traces in context to cut repeated work in agent loops. Same hybrid Gated DeltaNet and gated-attention design as Qwen3.5; 262K native context. Apache 2.0.
- Date
- Thursday, 16 April 2026
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| SWE-bench Verified | 73.4 (35B-A3B) / 77.2 (27B) 27B: MMLU-Pro 86.2, AIME 2026 94.1; 35B-A3B: GPQA 86.0, AIME 2026 92.7 | company |
35B-A3B released 2026-04-16 and 27B on 2026-04-22 per the QwenLM README (HF repos created 2026-04-15 and 2026-04-21). The proprietary Qwen3.6-Plus (2026-04-02) and Qwen3.6-Max-Preview (2026-04-14), plus Qwen3.5-Omni, are separate records.
Sources
- huggingface.co/Qwen/Qwen3.6-35B-A3B
- huggingface.co/Qwen/Qwen3.6-27B
- github.com/QwenLM/Qwen3.8
- www.alibabacloud.com/help/en/model-studio/newly-released-models
This record was checked against its sources on 6 October 2026. How we check