Qwen3.6-27B 3090 tuning myths spread

The hot take: this looks less like a fatal Qwen3.6 architecture flaw and more like local-inference cargo culting. The official release positions Qwen3.6-27B as a dense, practical coding model, while early community reports show it can run at high context on a single 3090 if the config is disciplined. Qwen’s April 22, 2026 release pitches Qwen3.6-27B as a dense 27B multimodal model meant to be easier to deploy than MoE alternatives, not as a weird edge-case architecture that inherently breaks llama.cpp caching. In the Reddit thread, experienced llama.cpp users directly reject the claim that KV cache cannot be reused, and say recent builds can reuse checkpoints correctly. The posted config is bloated: duplicated `-ngl`, quantized KV cache, `mmproj-F32`, 128K context, and large ubatch values all add memory pressure or slow prompt processing before any model-specific limitation enters the picture. Other same-day community posts report 64K to 127K context on single 3090-class cards and big speed gains from speculative decoding, which undercuts the idea that `graph splits = 34` proves unavoidable CPU spill or a universal Qwen3.6 bottleneck. The broader lesson for local AI infra: dense 27B models are now good enough that tuning mistakes, not raw model quality, are often the real blocker on consumer hardware.