anikifoss

/

DeepSeek-V3.1-HQ4_K

like 2

DeepSeek V3.1 with thinking mode is the first model that has acceptable vibe-coding performance for me. I was able to put together this game by providing LLM feedback and requesting changes over multiple prompts. Everything was made by the LLM: the game code, the enemies, the sounds, the dialogs. I tweaked some difficulty settings by hand, but only because the code was very long and it was faster than waiting.

This is also the first open source model to make chagnes over 3k loc without losing the plot! While the benchmarks don't suggest it, the performance is truly impressive.