hey everyone!
i’ve been working on getting large open weight models running well on a single dgx spark. i’ve had a decent bit of luck with deepseek v4 flash at 1M, and have now been working through creating an easy toggle to switch to glm 5.2 at 1M as well (this is still a work in progress, i only have preliminary results) .
i thought, perhaps, other people would be interested in this, so i have created a repo to hold everything that i have currently done, created a lot of instructions for agents, and (hopefully) configured it so that other agents can easily contribute.
what i’ve found so far is that, using codex sol @ medium and having it set a goal works decently well, and run it directly on your spark. you should, in theory, be able to point your agent directly at the repo and have it work. if it doesn’t, then i’d want to treat it as a bug/issue and open a PR to get it working.
please take a look and let me know what you think! if you have a spark and some extra codex usage and some time, i’d be grateful for the help! also, any and all feedback would also be helpful for how this could maybe work even better.