antirez Tests Domestic AI Models for Coding: DeepSeek vs GLM vs Kimi

antirez · x · 2026-08-01

Prominent developer antirez shared his hands-on experiences with recent mainstream AI models. He noted that DeepSeek is solid for local deployment and coding, but still trails GLM 5.2 and Kimi K3 in overall capability, acknowledging they target different categories.

In practical testing, he attempted to generate the game Frogger using the mxfp4 quantized version of DeepSeek via omp. It failed three consecutive times, producing games that were either too complex or ridden with minor bugs. However, DeepSeek excelled when used for web search and resolving HTML/CSS issues with custom plugins via Hermes Agent.

Original post →

More from coding & agent

coding & agent channel →