Qwen3.6-27B Outperforms 10× Larger Qwen3.5-397B on Coding Tasks

Qwen3.6-27B Outperforms 10× Larger Qwen3.5-397B on Coding Tasks

Alibaba's Qwen3.6-27B — a dense 27B model optimized for agentic coding tasks including planning, repo navigation, bug fixing, and tool use — outperforms the 397B parameter Qwen3.5-397B-A17B MoE on most serious coding benchmarks. The model supports dual think/no-think modes and multimodal reasoning. Available on HuggingFace (including FP8 variant), ModelScope, and Qwen Studio under Apache 2.0. The result challenges the assumption that parameter scale determines coding capability.

Why It Matters

A 27B model beating a 397B model on agentic coding validates that post-training quality, data curation, and task specialization are now the primary capability levers — a signal that matters for teams choosing between large and small model deployments.