Qwen3.6-27B Beats 35B on Coding Precision

The hot take is that local coding quality still rewards the smaller, denser model when the task demands correctness over raw generation speed. Qwen’s own release context reinforces that this is a serious coding-family benchmark, not just anecdotal speed-testing.

• –The result matches Qwen’s positioning for Qwen3.6-27B as a dense model aimed at flagship-level coding, with official benchmarks showing strong coding and agentic performance.
• –The 35B model’s higher TPS makes it better for interactive latency-sensitive workflows, but that speed advantage appears to come with weaker problem-solving on this prompt.
• –For coding primitives, precision matters more than throughput: one wrong abstraction or incomplete implementation can erase any benefit from faster token generation.
• –This is a useful reminder that “bigger” does not always mean “better” for local inference, especially when model architecture and training alignment differ.
• –The most practical takeaway is to choose by task: 27B for correctness and deeper reasoning, 35B for faster iteration loops and broader responsiveness.