pi v0.81.0: local model management with llama.cpp

pi v0.81.0 was published on July 21, 2026, and its headline feature is local model management. pi can now connect to a llama.cpp router, search and download models from Hugging Face, and explicitly load or unload them with live progress reporting. This makes running open-weight models locally a first-class workflow instead of a manual side setup.

Extensions also gained a major capability: they can now register complete pi-ai providers, including authentication, model refresh, filtering, and custom streaming. Previously extensions worked around the edges of the provider system; now a new provider can be shipped entirely as an extension.

The release further adds built-in Qwen Token Plan providers — international and China subscription variants with regional endpoints and API-key authentication — and expands usage accounting to cover tool calls, compaction, and branch summaries, giving users a more complete picture of where tokens go.

A same-day follow-up, v0.81.1, added deterministic, checksummed source archives to GitHub releases with instructions for rebuilding standalone binaries, plus more resilient retries for compaction and branch summaries when providers fail transiently.