Qwen3.6-27B sparks local tool-calling debate

The thread reads like a practical verdict on local agent models: if you care about tool reliability more than raw speed, Qwen still looks like the default answer.

• –OP wants a generalist model for housekeeping tasks, not coding, and is trying to balance decent reasoning with tolerable latency.
• –Several commenters argue Gemma 4 is weaker for agentic loops and function calling than Qwen at these sizes.
• –A few lighter-weight alternatives come up, including DFLASH and REAP, but they are framed more as speed compromises than clear upgrades.
• –The useful takeaway is that tool-calling quality is still a niche where "fast enough" often loses to "actually follows the loop."
• –For M4 Pro 48GB users, the real tradeoff is likely between Qwen3.6-27B-class density and MoE-style models like Qwen3.6-35B-A3B.