ios-ane-h18p: 6-bit k-means g8 static iOS export, AOT h18p neural-engine (replaces the 4-bit g32 export of 2026-09-15; device gate 3/4 exact, long answer diverges on a 0.17-margin step)
db8eb5b verified | { | |
| "metadata_version": "0.2", | |
| "kind": "llm", | |
| "name": "minicpm5_2b_minicpm5_pal6_g8_static", | |
| "assets": { | |
| "main": "minicpm5_2b_minicpm5_pal6_g8_static.h18p.aimodelc" | |
| }, | |
| "language": { | |
| "tokenizer": "openbmb/MiniCPM5-2B", | |
| "vocab_size": 130560, | |
| "max_context_length": 4096, | |
| "embedded_tokenizer": true, | |
| "function_map": { | |
| "main": [ | |
| "main" | |
| ] | |
| } | |
| }, | |
| "source": { | |
| "model_definition": "torch", | |
| "hf_model_id": "openbmb/MiniCPM5-2B" | |
| }, | |
| "compression": "minicpm5_pal6_g8", | |
| "compilation": { | |
| "date": "2026-09-16T10:55:25.976722+09:00", | |
| "targets": [ | |
| "h18p neural-engine (xcrun coreai-build compile --platform iOS --preferred-compute neural-engine --architecture h18p)" | |
| ] | |
| } | |
| } |