Qwen3.6-27B sparks local coding debate

Qwen’s current sweet spot looks less like raw parameter count and more like how much of the model is actually active at inference time.

• –On 24GB VRAM plus 64GB system RAM, a dense 27B model is the safer path for consistent speed and simpler offload behavior.
• –Qwen3.5-122B-A10B is already available, and its 10B active-parameter design can make it surprisingly practical on RAM-heavy rigs.
• –For coding, "bigger" does not scale linearly into "better"; quantization, context length, and runtime stack matter as much as model size.
• –Qwen3.6-27B is notable because it shows a dense model can compete with much larger releases on agentic coding tasks.