I did a few quick test runs with frameworks that I usually run against K3 / GLM 5.2 and it looks promising at first glance. Thinking blocks seem to have changed to be much more condensed by not spitting out full sentences, e.g.:

Should enough info. Could run actual tests with package standalone dependencies to verify behavior and regression noted, rather than inference. Acceptance just report no need execute but robust.

Also I noticed much more usage of rg instead of grep (using exactly the same system prompt between K3 and K3.1)