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

Alibaba’s Qwen team has officially released Qwen3.6-27B, a dense open-weight model that marks a significant milestone in the Qwen3.6 series. As the first fully dense model in this family, it is designed to prioritize real-world coding utility and stability over mere benchmark optimization. Despite its 27-billion-parameter size, the model demonstrates superior performance on agentic coding tasks, outperforming both the Qwen3.6-35B-A3B sparse Mixture-of-Experts (MoE) model and the much larger Qwen3.5-397B-A17B MoE.

Advancements in Agentic Coding

The model is specifically optimized for complex software engineering workflows, including repository-level reasoning and frontend development. In internal testing on QwenWebBench, Qwen3.6-27B achieved a score of 1487, significantly outpacing its predecessor, Qwen3.5-27B, which scored 1068. Its capabilities are further evidenced by a 77.2 score on SWE-bench Verified and a 59.3 score on Terminal-Bench 2.0, a result that matches the performance of Claude 4.5 Opus.
A standout feature of this release is the Thinking Preservation mechanism. By enabling this option, users can retain and leverage chain-of-thought reasoning traces from historical messages across a conversation. This approach allows the model to carry forward context rather than re-deriving it in iterative agent workflows, which reduces redundant token consumption and improves KV cache utilization.

Hybrid Architecture and Technical Specifications

Qwen3.6-27B utilizes a sophisticated hybrid architecture across its 64 layers. It employs a repeating pattern where three out of every four sublayers utilize Gated DeltaNet—a form of linear attention that provides O(n) complexity—while the fourth sublayer uses standard Gated Attention. This design choice enhances memory efficiency and speed, particularly for long-context tasks. The model also incorporates Multi-Token Prediction, which facilitates speculative decoding to improve throughput during inference.
The model supports a native context window of 262,144 tokens, which can be extended up to 1,010,000 tokens using YaRN scaling. It is natively multimodal, supporting text, image, and video inputs. The Qwen team has released two weight variants on the Hugging Face Hub: a 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.

Performance and Availability

Beyond its coding prowess, Qwen3.6-27B shows strong results in general reasoning and multimodal tasks. It achieved 87.8 on GPQA Diamond and 94.1 on AIME26. Its vision-language capabilities remain robust, scoring 87.7 on VideoMME and 70.3 on the AndroidWorld visual agent benchmark. Licensed under Apache 2.0, the model is now available for developers looking to integrate high-performance, dense-model capabilities into their own agentic systems.