icon
Programmer's Diary

Weekly Coding β€’ 2026-08-17 β€’ From Manual Sync to Automated Flow: Exploring Tygo, Qwen 3.8, and Smarter Development Workflows.

This text was automatically translated to English language by LLM tool.

Last week, I spent 11 hours on my coding adventures in my free time.

Personal Website: Code Improvements 🌐

I implemented code generation using the Tygo tool.

My client and server are written in different programming languages: TypeScript and Golang. Previously, the code for structures passed between the client and server had to be manually written in both languages. Added a new field? You had to write it in two files: .go + .ts.

Now, Golang is the source of truth, and the TypeScript files are generated automatically during every build. An example is shown in the screenshot.

Previously, I thought it would be convenient to integrate a full-scale large REST framework into the project so that not only the common structures but also the entire REST client would be generated automatically. Instead, I’ve implemented this intermediate solution. Was it worth it? πŸ€”πŸ’­ Or would it be easier to just switch to a large REST framework right away? The main advantage of the intermediate approach is that there are fewer dependencies in the project. Yes, a code generator is also a dependency, but it’s small. It sort of sits on the side rather than permeating the entire project.

Local AI: Qwen 3.8 27B ❄️

A new model for local LLM coding has been released. It’s the Qwen 3.8 27B. I tried out this model over the weekend and used it to implement a TypeScript code generator for my project. The results are as follows: the model works surprisingly well. As a result, here is my current top 3 models:

  • ❄️ Qwen 3.8 27B πŸ† best code; good logic
  • 🐀 Ornith 1.0 35B a3b πŸ† good code; shaky logic due to Mixture of Experts
  • ✨ Gemma 4 31B QAT πŸ† average code; smart logic

My thoughts on local LLMs:

  1. πŸ“‰ If given a massive task with complex legacy code, a local model will get stuck. And I’ll get stuck too (
  2. πŸ’Ή If you break the task down into small sub-tasks, a local LLM works as an accelerator for manual coding. That is exactly how I use local LLMs right now.

I have an idea to start a new project in my free time, but I haven’t decided yet because having enough time for rest and health is more important to me.