Qwen 3.8 27B Coder390 for AI Pod

Independent LPK package: cloud.lazycat.aipod.qwen38-27b-coder390.

This repository contains the verified NVFP4 W4A4 target and FP8 DFlash2 snapshots, a separately published model-free Thor runtime, and the minimal LPK controller release. No files or release IDs of other Qwen 27B projects are updated.

Upstream model: nerkyor Coder390, fixed revision bae2cbd736952a21006feca2630f5715016250c2.

Target hardware: NVIDIA Thor SM110, 128 GB unified memory. Default context: 32768 tokens; scheduler limit: 8 requests. Performance reports identify their exact images and test conditions.

Validated LPK release: 0.1.2

huggingface · modelscope

LPK SHA256: a732c042b3fd1cd549e539942c0f490332e2c04e10078791116ffd20012cb974; size: 171008 bytes.

The 0.1.0 and 0.1.1 LPK files are incomplete release candidates. Use 0.1.2, which passed the installed source matrix on 192.168.1.160. The model snapshot and validated inference runtime retain their separate immutable identities.

The installation tests reused fully SHA256-verified model bytes in an independent device namespace; both public model snapshots were separately downloaded anonymously and fully verified. The final runtime was imported from an absent image state for both sources.

See release-summary-0.1.2.html, release-summary-0.1.2.json, and the profiling/validation evidence. Desktop/mobile entry and the real model-store card were tested; full LazyCat portal session login was not covered.