Qwen3.6-27B
Alibaba's first dense open-weight Qwen3.6 — 27B beats 397B MoE on agentic coding via hybrid Gated DeltaNet + Gated Attention
Qwen3.6-27B
Qwen • April 2026
Training Data
Up to early 2026
Qwen3.6-27B
April 2026
Parameters
27 billion (dense)
Training Method
Hybrid Gated DeltaNet linear + Gated Attention (3:1), Multi-Token Prediction
Context Window
262,144 tokens (1M with YaRN)
Knowledge Cutoff
Not disclosed
Key Features
Apache 2.0 Open Weights • Thinking Preservation • Multi-Token Prediction • Native Multimodal (text/image/video)
Capabilities
Coding: Outstanding
Agentic: Outstanding
Multimodal: Excellent
What's New in This Version
Beats 397B MoE peers on SWE-bench Pro (53.5); matches Claude 4.5 Opus on Terminal-Bench 2.0 (59.3) at 14× fewer parameters; SWE-bench Verified 77.2
Alibaba's first dense open-weight Qwen3.6 — 27B beats 397B MoE on agentic coding via hybrid Gated DeltaNet + Gated Attention
What's New in This Version
Beats 397B MoE peers on SWE-bench Pro (53.5); matches Claude 4.5 Opus on Terminal-Bench 2.0 (59.3) at 14× fewer parameters; SWE-bench Verified 77.2
Technical Specifications
Key Features
Capabilities
Other Qwen Models
Explore more models from Qwen
Qwen3.6-Max-Preview
Alibaba's most powerful Qwen yet — preview flagship topping six major coding benchmarks
Qwen3.6-Plus
Alibaba's flagship agentic AI model with hybrid linear attention, always-on reasoning, and autonomous multi-step coding workflows
Qwen3.5-Plus
Alibaba's hosted flagship combining hybrid linear-attention MoE with native multimodal understanding for agentic workflows across 201 languages
