Introduction You’re debugging a frontend engineering issue — the page is behaving abnormally. You ask an AI to open the page with a browser tool and check the console for errors.
The AI opens the page, scans around, and tells you: The console is clean, no errors whatsoever.
You’re skeptical. You open Chrome DevTools yourself — three bright red errors are staring you in the face, the page has already crashed into a white screen. The AI visited the exact same page using a Headless browser, so why did it catch nothing?
One Month In It’s been nearly a month since the previous article Switched My Blog Theme: From Hugo NexT to Self-Written Zhi. During this month, the theme has been running stably without major issues. Replacing the NexT theme was the right decision — although there was some initial adjustment, the experience is now significantly better.
After a month of use, the theme’s stability has exceeded expectations. My initial concerns — whether pure Hugo Pipes without build tools could support complex requirements — were proven unfounded. Daily maintenance has become very simple; modifying a feature no longer requires digging through deeply nested SCSS files — one CSS file gets the job done.
The previous article covered the initial configuration setup. This one documents the adjustments after two weeks of running: expanding from single vendor to a four-tier model pool, adding fallback chains, hitting the GLM-4.5-air trap of analyzing without writing code.
This post covers: fallback strategy design, complete free model pool inventory and analysis, concurrency control configuration, and the decision process for GLM-4.5-air replacement.
After the previous article’s initial configuration, I ran it for two weeks — all the issues that needed fixing surfaced.
Why Switch This blog previously used Hugo NexT, forked for custom modifications. NexT itself is a feature-rich theme, but when it comes to “customizing things yourself,” the experience wasn’t great.
The issues boiled down to a few things:
SCSS nesting hell. 101 SCSS files, three levels of directory nesting. _common/components/post, _common/components/third-party, _common/outline/sidebar… To change a style, you first had to figure out which file it was in, where variables were defined, and which scheme was overriding it. Not that it couldn’t be done, but each change meant half an hour of hunting.
Why Bother When it comes to writing code with AI, the gap between single-model and multi-model approaches keeps widening. No matter how strong a single model is, it can’t compete with a team of specialized models working in parallel.
Oh My OpenCode (OmO for short) is a multi-model orchestration plugin in the OpenCode ecosystem, with 11 Agents each having distinct responsibilities and 48 Hooks spanning the entire lifecycle. Zhipu’s Coding Plan provides access to the full GLM model series. Combining the two allows you to assign different models by role — strong coders for coding, strong reasoners for reasoning, free models for busywork.
Overview This article compares the resource requirements and usage costs of three major domestic LLMs, helping developers choose the right solution for their scenarios.
Model Vendor Architecture Minimum Deployable VRAM API Available GLM-5 Zhipu AI Dense (multiple versions) 24GB (8B) ✅ Kimi K2.5 Moonshot AI MoE (undisclosed) 24GB (lightweight) ✅ MiniMax M2.7 MiniMax MoE 230B Not yet open-sourced ✅ GLM-5 (Zhipu AI) Versions & Hardware Requirements GLM-5 offers 4 parameter versions, making it the widest-coverage domestic LLM currently available.