Skip to content

Hive LLM

Fabrizio Degni edited this page Nov 15, 2026 · 2 revisions

Hive LLM engine

Implementation and setup · Actual pilot and graphs

Hive owns the request and lease lifecycle. Enrolled workers execute a complete resident model. The real pilot used a separate Linux native runtime co-located with the hub, without Mac inference or relay. Mobile adapters compile; their physical LLM qualification remains pending. No pooled model memory, general speedup or energy saving is claimed.

All registered devices now remain visible, including forwarded/non-ready entries. Qwen2.5 72B (145.4 GB) and Kimi K2 (1.029 TB) are metadata inspection targets only. They are not admitted for execution. Physical OPPO results and capacity boundary.

Clone this wiki locally