Qwen3.5-397B-A17B and Qwen3.5-Plus
Open 397B-A17B native vision-language MoE with Gated DeltaNet linear attention, 201 languages and 262K context; hosted Qwen3.5-Plus offers 1M.
Hybrid of Gated DeltaNet (linear attention) and sparse MoE, 512 experts with 11 active, early-fusion multimodal training, and language coverage up from 119 to 201. Alibaba says it matches Qwen3-Max with 8.6x to 19x faster decoding. Apache 2.0.
- Date
- Monday, 16 February 2026
- Lab
- Alibaba (Qwen)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| MMLU-Pro / GPQA | 87.8 / 88.4 SWE-bench Verified 76.4 on the HF card; Alibaba's blog lists 80.0 for the hosted setup; MMMU 85.0, MathVision 88.6 | company |
Open weights appeared 2026-02-16 (HF repo creation 04:55 UTC); Alibaba Cloud's blog repost is dated 2026-02-17 and Model Studio lists the hosted qwen3.5-plus on 2026-02-15 and qwen3.5-flash on 2026-02-23. The hosted Plus context is 1M by default; the open model is 262,144 native, extensible to about 1.01M. MMLU-Pro 87.8, GPQA 88.4 and SWE-bench Verified 76.4 match the HF card; Alibaba's blog lists 80.0 for the hosted setup. All numbers are company-reported.
Sources
- www.alibabacloud.com/blog/602894
- huggingface.co/Qwen/Qwen3.5-397B-A17B
- huggingface.co/api/models?author=Qwen&sort=createdAt&direction=-1&limit=60
- www.alibabacloud.com/help/en/model-studio/newly-released-models
This record was checked against its sources on 6 October 2026. How we check