Anyone tried to challenge the model?
I decided to try and create a chess game in html and so far it's not easy at all. The model seems very stupid at times. I'm doing everything step by step because it seems impossible in one go (btw gemini 3 pro did it with just one promt, even the bot opponent is working).
But this model...first it created just the chess board and nothing else as i promted, but the dark and light squares were swapped, not correct. So i told it to swap the squares, light with dark. And it failed! It's an easy fix, i just did it myself then... just these two lines of code
--board-light: #b58863; /* Light Wood /
--board-dark: #594034; / Dark Wood */
swap the colors and done, but this model overthinking like crazy, and even broke the whole board.
Now i'm trying to add the chess pieces, trying... man this model just doesn't know chess it seems...
Maybe anyone willing to try this too? 😀 and tell us if it's possible at all.
I'm using llama cli, UD-Q4_K_XL.gguf and settings seems to work well --multiline-input --jinja --temp 0.7 --top-p 1.0 --min-p 0.01 . But this task seems impossible.
"gemini 3 pro did it with just one promt"
Your observation is correct that GLM-4.7-lash is not competitive with Gemini Pro.
Personally, I dont rely on small models to get code right, but they can suggest code that I might choose to use after modification.
Actually, i managed to create a chess game against bot in html using this model. It just takes time...step-by-step, only using llama.cpp. So this model impressed me. I don't mind spending some time improving code and fixing bugs :) At least it works well
Later i created texas holdem poker against 3 bots, it's not bad, just some gameplay bugs that need to be fixed.