15 November 2026 · Experimental complete-model placement

Hive LLM. Actual execution.

Hive now owns requests, model admission, worker leases, output, cancellation and terminal receipts. Enrolled workers execute their model directly; a Mac is not required for inference, verification or relay.

19 / 18 / 2

19 ordinary Linux requests completed. All 18 matched native-reference comparisons preserved token IDs. Two deliberately injected failures confirmed cancellation and lease expiry, with late output rejected.

Scope: one Linux VPS, separate hub and worker processes, Qwen2.5 0.5B Q4_K_M, one CPU thread, native CPU quota 60%. This is not a mobile/WAN benchmark or shared-model-memory implementation. No general acceleration or energy saving is claimed.

Transport observations

Six observations per mode. Median blocking control wait was 0.292 s with new connections and 0.240 s with reuse. Median paired Hive/native completion ratios were 1.064 and 1.076. Less control wait did not establish whole-request acceleration.

Android 0.2.28/build 32 and unsigned iOS compiled. Subsequent OPPO testing completed six bounded Hive requests, reported below. Sustained mobile stability, iOS execution and multi-phone execution remain unqualified.

Measured charts

Complete paired timings
Complete paired timings · Linux CPU / Qwen2.5 0.5B
Exclusive wall-time phases
Exclusive wall-time phases · Linux CPU / Qwen2.5 0.5B
All transport observations
All transport observations · Linux CPU / Qwen2.5 0.5B

Model download: 491,400,032 bytes streamed over actual authenticated Hive HTTPS in 8.581 s, with matching SHA-256. Co-located Linux recipient; acquisition is excluded from request timing.

Read the full report, raw receipts and reproduction ↗

Install updated companion ↗

Physical OPPO execution and the capacity gap

Six real phone requests matched their six native references. Median paired Hive/native duration was 3.722 in this fixed-order diagnostic session. A controller interruption and terminal instrumentation socket error remain unresolved. This is not multi-device model scaling.

The device pool now shows every registered entry and its readiness/enrollment reason. Qwen2.5 72B (145.4 GB) and Kimi K2 (1.029 TB) are inspectable checkpoint metadata only. The required-partition executor on Hive is still missing; no large weights were downloaded.

Physical OPPO matched native and Hive timing

Full phone evidence and remaining capacity requirements ↗