Qwen3.6-27B on 2x3090s trails 35B-A3B

The hot take is that this reads more like a serving-stack and quantization problem than a model-quality problem.

• –The post is not a benchmark claim; it is a troubleshooting thread with one reply asking for more configuration details.
• –The user specifically mentions vLLM and llama.cpp failures, which points to inference setup, not just prompt quality.
• –Qwen’s own release materials frame Qwen3.6-27B as a dense model optimized for agentic coding and long-context use, so bad tool settings can easily mask its strengths.
• –The comparison target, Qwen3.6-35B-A3B, is an MoE model; depending on quantization and runtime, the smaller dense model can feel better or worse in real agent workflows.
• –“Fails on writing big files” suggests context management, output truncation, or agent orchestration issues, not necessarily raw reasoning weakness.