Personal Website 🌐
Last week, I spent 16 hours coding my personal website. I managed to tackle the following tasks:
-
Switching from LM Studio to Ollama for translating text into
foreign languages and headers
- Main reason for the move: Ollama doesn’t consume VRAM.
-
Switching from the Aya Expanse 8B model to Gemma 4 12B QAT
- The new model is more modern and powerful. The previous model was released in 2024.
-
Switching from HTML to Markdown for storing my blog posts
- LLMs now handle my texts better because they are more optimized for Markdown.
- Main reason for the move: I frequently had issues with text generation before—the model would mess up and output mangled HTML, or it would try to convert HTML to Markdown even though I hadn’t asked it to do so.
Local LLM 🤖
Once again, I tried to set up a local LLM for coding. This time, I tried the GPT OSS 20B model. The model is functional but somewhat weak and prone to errors. For example, it generated a Regex for me but forgot to add *\* before ?—a classic amateur mistake. As a result, the Regex doesn’t work. However, the speed is satisfying: I get enough tokens per second because the entire model fits in VRAM 👌
I am using the LM Studio + OpenCode combination.
The weekend came and went 🗓️🫡
On Saturday and Sunday, I racked up quite a lot of hours in my tracker. 8 + 8 = 16; that means I spent practically a full workday over the weekend. However, the time passed without any stress because I worked on my project tasks very slowly and relaxed, taking plenty of breaks.
