Repository navigation
NodeLift 0.2.28 / build 32: direct Hive LLM workers
Pre-releaseNodeLift 0.2.28 / build 32
Hive-owned complete-model requests and direct mobile workers
- Add persistent Hive-owned LLM requests, model hash admission, bounded greedy generation, fenced worker leases, incremental output, cancellation and immutable result replay. Device permission and operating windows remain separate; the one-request exception is explicit.
- Add direct Android and iOS complete-model worker adapters and authenticated model streaming from the private Hive catalog. Model execution stays on the selected phone; the Mac is not an inference host or request relay.
- Add the LLM tab to Hive with model/node selection, output and separate queue, first-token, prefill, generation, control-wait and service timings. English, Simplified Chinese and Italian messages are included.
Validation: 475 host tests, 55 Android JVM tests, APK/lint and unsigned iOS ARM64 compilation passed. A separate unprivileged native Linux worker co-located with Hive completed 19 ordinary requests and 18 matched native-reference comparisons. Two injected failures preserved cancellation/expiry and rejected late output. Authenticated model streaming transferred 491,400,032 hash-matching bytes. Android installed on OPPO, but its locked screen prevented physical LLM qualification; iOS physical execution and live owner click-through remain pending.
Limits: Complete-model placement only: no model splitting or shared model memory. One operator-provisioned default mobile model, at most 1 GiB, context up to 4096 and 256 output tokens, greedy text only. Results and device phase times are authenticated worker reports, not independent semantic attestation. Hive times use server acceptance and receipt boundaries. No acceleration or energy saving is claimed. iOS archives require personal signing and physical qualification.
Android: API 26+ / ARM64, same experimental debug signing identity as 0.2.27. Install over the existing app to preserve enrollment. In Hive mode, prepare the registered model, enable participation and resume the session.
iOS: both compiled archives are unsigned, with identical app payloads. Personal Apple signing/provisioning is required; neither installs directly from Safari. No TestFlight/App Store release.