So re: The coding agent project.
I'm still on schedule for a Monday completion.
Mac and Linux are supported cleanly, but Samuel and I are still working with capability and inference issues on the M3.
We might have a workable runtime for mac by release time, we might not. It's still in flight.
The system runs beautifully with GLM and Gemma4, and Qwen 3.5 models.
Gemma4-26b is really iffy, but I'm going to do more investigation. Can't seem to make it through the whole workflow without intervention. Planning is weak, multi-step is also weaker than Qwen 3.5 9b.
But... it's entirely possible I'm running a lossy 4bit Unsloth version that drifts too much on my 4070. Tool system differences are fully accounted for, which is good, but I need some clearer indications that I'm running a reliable installation with it.
Kimi K2 should also be supported, along with most other models, unless your provider does weird things with the tool channel.
I've accounted for all the common edge cases there, but depending on things like the Jinja templates and whatnot, there's no way to anticipate every variation someone could use.
Still feel pretty good about it.
The tool is this sort of new theory on agentic labor. And I think people are going to like it.