Qwen Launches Qwen3.6-27B Open-Source Model with Top Coding Performance

Details

• Qwen released Qwen3.6-27B, a 27B parameter dense open-source model excelling in coding, surpassing the larger Qwen3.5-397B-A17B (397B total, 17B active) on key benchmarks.
• Coding scores: SWE-bench Verified (77.2 vs 76.2), SWE-bench Pro (53.5 vs 50.9), Terminal-Bench 2.0 (59.3 vs 52.5).
• Natively multimodal, supports vision-language tasks with images and video in unified checkpoint, matching Qwen3.6-35B-A3B capabilities for reasoning and document understanding.
• Delivers flagship-level agentic coding despite smaller size, emphasizing efficiency over scale.
• Available openly, building on Qwen's series of accessible high-performance models like prior 27B and MoE variants.

Impact

Qwen3.6-27B demonstrates dense models can rival much larger MoE architectures like its 15x bigger sibling on coding benchmarks, pressuring rivals such as DeepSeek and Llama series that prioritize scale. By matching GPT-5 mini-level SWE-bench scores at 27B fully open-weights, it lowers barriers for developers building cost-effective agents, widening access to frontier coding and multimodal capabilities without proprietary restrictions. This advances open-source competition, narrowing gaps with closed models from OpenAI and Anthropic in practical inference efficiency.