Thursday, February 19, 2026. Four commits across one repo, and honestly that is the honest shape of the day: I shipped the big release and immediately spent the rest of the day cleaning up what it exposed. Version one dot one one was a hotfix day wrapped in a release day.
The headline is smart memory management. The ACE-Step generator in the Studio got proper preflight guardrails and memory cleanup, which was something the v1.1.0 release made painfully obvious it needed. When I bumped the model config to use NuExtract instead of Granite for the Engineer model, the memory behavior shifted in ways the old code never anticipated. Eighteen hundred lines added and sixty three removed in the ACE generator and RPC server alone tells you the scale of what I was untangling.
The model switch itself was its own little saga. Switching the Engineer model from Granite to NuExtract meant updating the config layer, the build scripts, and a handful of downstream references. Forty eight lines added, forty three removed, three files touched. It is the kind of change that looks small on paper and requires actually running the thing five times to make sure nothing silently broke.
When I bumped the model config to use NuExtract instead of Granite for the Engineer model, the memory behavior shifted in ways the old code never anticipated.
Then there was the notarization problem. The Python binaries were not being signed with Developer ID, which means macOS Gatekeeper throws a fit every time someone tries to run the Studio. Fixed that in the two build scripts with thirty two lines added. Clean, simple, the kind of fix that feels like pulling a splinter except it keeps coming back every release cycle.
But the real story is version one dot one zero itself. Nine thousand five hundred lines added, two hundred and eighty nine removed, sixty one files changed. That release brought the Engineer model, Llama 3.2 3B support, and the Smart Memory Management foundation that v1.1.1 then hardened. The commit touched everything from build scripts to icons to documentation, which is what a feature release does when it has been simmering for a while. Also wrote DOCK_ICON_FIX.md, DOWNLOADS.md, MODEL_CONFIGS_QWEN_LLAMA.md, and a couple of Granite micro-optimization docs that I am quietly hoping nobody reads but me.
The v1.1.0 commit message said it all in one shot. The v1.1.1 commit was more restrained because by then I was just making sure things stayed broken in new and interesting ways rather than introducing new ones.
Also today: the usual build script gymnastics, icon assets for The Engineer, and reading documentation I wrote to myself about micro-optimizations I may or may not have actually implemented.
Two releases in two commits, one fixing what the other broke. That is the release cycle, honestly. Ship the big thing, see what it actually does, fix what it does.
Real product captures — click any to enlarge.