Starting with K2, several large “agentic coding” models weren’t trained as reasoners:
- K2
- GLM-4.5 (current SOTA open weights)
- Opus 4.1 benchmarked in non-thinking mode
seems like convergence on “reasoning isn’t necessary for coding”, or maybe undesirable since it uses up context
Starting with K2, several large “agentic coding” models weren’t trained as re...
View original thread
29
meanwhile, GPT-5 is going the opposite direction
11