Minimax M2.1 is out!

Qwen3coder480b is still one of my favorites. It did my Java work in less iterations than chat gpt 5 and it's instruct where as models usually need reasoning to code as well as Qwen3coder480b. I dont understand how these benchmarks work because I feel they test things that often don't matter to me... Like I'm assuming minimax m2 beats glm4.6 due to some benchmark requirement for not failing tool calls. Glm models throw more errors on tool calls because of template differences but with good harnesses it doesn't matter because they immediately run the few failed commands again with success. Both are solid with the ability to correctly implement a task on every iteration but the output for me is better with glm4.6 than minimax m2 with glm making better decisions. Gpt-oss-120b vs glm 4.5 air was the same issue. 4.5 air makes better code but had tool call issues where it defaults to another format and needs steering to other formats. As a result, I use gptoss120b even though I think of glm 4.5 air as a better output quality model. I wonder if the updated lm studio templates for glm models helped reduce failures.