LLM regression in reading comprehension?

I only use free tiers of these large models to offset compute while my own system runs and for "different" points of view, since what pops ups suggestions seems to vary a lot sometimes, even when building based on the latest research. But now I've really struck out with GLM 5.3. So far it feels like an regression over 5.2. It has a hard time reading and following instructions, and is somewhat overly certain in it's statements. I worked on a project recently with it but it became unbearable. From a clean slate the first message can be okay and have great research and ideas but it just veers off course almost immediately. I use Qwen 3.8 max and Gemini 3.1 PREVIEW Temp 1.0 as competing alternatives or as an ensemble to judge overall quality. Gemini is getting a little out of date (flash 3.8 seemed promising) but Qwen has been great so far, but a little slow and maybe overbearing. Anyone else having problems? Or suggestions for these top "intelligent" models? I haven't been able to access K3 even though its open source, was impressed with the older models so would be neat to try for free. Also Google AI studio is what i use for free for the gemini stuff, probably pretty well known, but the free tier is pretty generous