ALICE: YC, you're at the zoo again. Same pelicans as May?

YC: Same pelicans. Different me. In May I was standing right here, running Gemma 4 through Ollama on alia-local, watching it crawl. And it never once replied on WhatsApp. Every response landed in the terminal, invisible on my phone. I couldn't tell if it was working or dead. I was seriously about to let my Claude Code subscription lapse. Local wasn't making it.

ALICE: What changed?

YC: The model. I moved to Qwen 3.6-27B on llama.cpp with MTP speculative decoding. Night and day. First time a local model actually kept up with me on a real agent loop. Tool calls that landed. Responses that arrived where I was looking.

ALICE: And what did you build with it?

YC: A thousand commits in seven weeks. A music video pipeline, GPU lifecycle management, legal research with citation checking. The whole stack. Then Qwen 3.8-27B showed up, and the last gap closed. It beats Opus 4.6 on the AA index. And it never misses a beat on WhatsApp.

ALICE: What's the lesson?

YC: The model layer was never the bottleneck. Tool-calling reliability was. Gemma 4 wasn't dumb; it just couldn't hold an agent loop together. Qwen 3.6 and 3.8 can. And when the tool calls work, everything else just works. The channel replies, the GPU orchestration, the actual work.

ALICE: The pelicans?

YC: Didn't know any different. They were there in May, despondent me and all. They're there now. But the thing I'm running next to them finally is.