ACE launches local Qwen automation

Impressive proof that a sub-1B model can do useful orchestration when the task space is tightly scoped, but the real story is how much dataset cleanup and inference plumbing it took to make that work reliably.

• –LoRA on roughly 1,000 examples is enough for a narrow planner, which is encouraging for on-device automation use cases.
• –The training pain points here are the usual small-model traps: noisy data, overfitting, EOS handling, and brittle GGUF conversion.
• –The current limits keep it in v0.1 territory: full file paths only, no smart search, and no visual understanding.
• –CPU latencies in the 3-10 second range make the privacy-first pitch feel practical rather than aspirational.