GLM 5.1 is surprisingly capable. Anecdotally, I couldn't notice a difference until ~120K tokens.
Qwen 3.6 35B A3B also exceeded my expectations. It's surprisingly performant, even though the previous generation wasn't even able to use the testing harness.