A user shared their experience with several open-source language models focused on coding, including Qwen 3.5 (in the 122B, 35B, and 27B versions), GPT-OSS 120B, StepFun 3.5, MiniMax M2.5, Qwen Coder 80B, and Super Nemotron 120B.

Qwen 3.5 397B: The Best for Local Coding

The user found that Qwen 3.5 397B stands out for its knowledge and ability to produce bug-free code. Despite being slower than the other tested models, the time saved due to the reduced need for corrections and its ability to provide concise answers makes it preferable.

Quantization and Hardware Requirements

A notable aspect is the use of IQ2_XS quantization, which reduces the model size to just 123GB. Other models were used with less aggressive quantizations such as IQ4_XS (StepFun 3.5, MiniMax M2.5) or Q6_K (Qwen 3.5 122b/35b/27b, Qwen Coder 80b, Super Nemotron 120b).