Open models may still struggle inside coding agents
The claim that are only about four months behind closed commercial models is being questioned through a coding experience. can still run into odd problems that rarely appear with commercial coding tools.
In one test, GLM5.2 was asked to create a setup for with Gemma, but it kept thinking for too long and repeating itself. The reasoning level was set to high, and the run was eventually cancelled.
Cursor was compared as a tool that can use a browser in the background and keep moving on similar tasks, while Zoo Code did not handle the same situation as smoothly. The open question is how often people compare with closed ones, and which they use, such as Zoo Code, Opencode, or Cline.
Key points
- The item challenges the idea that are only four months behind .
- GLM5.2 struggled with a coding task and repeated itself.
- High reasoning settings may increase token use when the model gets stuck.
- Cursor was described as smoother because it can use a browser in the background.
- Zoo Code, Opencode, and Cline are mentioned as possible choices.