Alibaba Releases Qwen3.6-27B: Dense Model Outperforms 397B MoE

Alibaba’s Qwen team has officially released Qwen3.6-27B, a dense open-weight model that represents a significant leap in capabilities for coding agents. As the first dense variant in the Qwen3.6 family, this 27-billion-parameter model demonstrates performance that surpasses both the sparse Qwen3.6-35B-A3B and the much larger Qwen3.5-397B-A17B Mixture-of-Experts (MoE) model on several key benchmarks. Released under an Apache 2.0 license, the model is designed to prioritize real-world utility and stability based on community feedback.

Advancements in Agentic Coding

Qwen3.6-27B is specifically optimized for repository-level reasoning and frontend workflows, enabling it to navigate complex file structures and maintain consistency across multiple files. On the SWE-bench Verified benchmark, the model achieves a score of 77.2, placing it in direct competition with Claude 4.5 Opus. Furthermore, it reaches a score of 59.3 on Terminal-Bench 2.0, matching the performance of Claude 4.5 Opus, and records a 1487 on the internal QwenWebBench, significantly outperforming its predecessor, Qwen3.5-27B.
The model also introduces a novel Thinking Preservation mechanism. While most large language models discard chain-of-thought reasoning after a single turn, Qwen3.6-27B allows users to retain and leverage thinking traces from historical messages. By enabling this feature via the API, developers can reduce redundant token consumption and improve KV cache utilization, which is particularly beneficial for iterative agent workflows.

Hybrid Architecture and Technical Specifications

Under the hood, Qwen3.6-27B utilizes a hybrid architecture that blends Gated DeltaNet linear attention with traditional self-attention. The model features 64 layers, where three out of every four sublayers employ Gated DeltaNet to achieve linear complexity, while every fourth sublayer uses standard Gated Attention. This configuration, combined with Multi-Token Prediction (MTP), supports speculative decoding to improve inference throughput.
The model is natively multimodal, supporting text, image, and video inputs. It comes with a native context window of 262,144 tokens, which can be extended up to 1,010,000 tokens using YaRN scaling. The Qwen team has made two weight variants available on the Hugging Face Hub: the standard BF16 version and a version utilizing fine-grained FP8 quantization with a block size of 128. Both variants are compatible with SGLang, vLLM, KTransformers, and Hugging Face Transformers.