Models / Qwen
Qwen3.6-27B
Strengths
Flagship-level coding in a 27B dense footprint. SWE-Bench Verified 77.2%, Terminal-Bench 2.0 59.3% (matches Claude 4.5 Opus). 262K native context, multimodal, Apache 2.0.
Weaknesses
27B dense weights need a 24GB GPU at Q4; bf16 wants 80GB+ - no MoE active-param shortcut. English / Chinese first; other languages quieter than Qwen3.5.
Qwen3.6-27B is the dense model that closed the gap with closed-source coding agents. SWE-Bench Verified at 77.2 puts it within four points of Claude Opus 4.6, and Terminal-Bench 2.0 lands exactly where Claude 4.5 Opus does - on a 27B dense checkpoint that fits on a single high-end consumer GPU.
The "thinking preservation" trick is the architecture detail to know: reasoning context from earlier turns can be retained across iterative edits, which compounds noticeably on long agent loops. For teams who can't host the 80B Coder-Next weights, this is the new default agentic-coding pick under permissive license.
When to pick it
- Agentic coding on a 24-48GB GPU. Frontier-class quality, Apache 2.0.
- Repo-scale tasks where 262K native context matters.
- You want one dense model for code, agents, and image-aware reasoning.
When to skip it
- 8-16GB hardware ceiling - drop to Qwen3.5-9B or Phi-4-mini.
- Pure latency-sensitive inference where MoE active-param speed wins - Qwen3-Coder-Next runs faster per token if you can host it.