🤗 mlboydaisuke/qwen3-0.6b-CoreAI-official · Apache-2.0 · source Qwen/Qwen3-0.6B
This is not a zoo port: every bundle in the repo is Apple’s own coreai.llm.export qwen3-0.6b
recipe, unmodified (the official/ line — see ../../official/README.md and
../../knowledge/apple-models-bench.md for the bench matrix).
This directory exists because the repo gained a Neural Engine bundle on 2026-09-15 that the zoo
gated on the phone, and a gated bundle needs a recipe entry and a transcript like any other.
| subtree | what | size | provenance |
|---|---|---|---|
macos/ |
macOS dynamic, int4 per-block (Apple’s macOS 4bit preset) |
335 MB | 2026-06-11 export, re-uploaded 2026-09-08 |
macos-26-export/ |
same recipe, macOS-26-era artifact (2.2× faster lowering; cannot be re-created) | 360 MB | 2026-06-11 |
ios/ |
iOS static IR, Apple’s iOS preset for qwen3-0.6b = models/qwen3/qwen3_0_6b_mixed_4bit_8bit.yaml (k-means 4-bit, per-grouped-channel group 8 on the body; layers 0/2/4/7/15/16 at 8-bit per-tensor; embeddings int8 by the iOS exporter), ctx 4096, static graphs prompt_opt/extend × {256, 512, 1024, 2048, 4096} × {8, 16, 64} + load_embeddings |
440 MB | 2026-06-11 export (coreai-torch 0.4.0), re-serialized 2026-07-21 by strip_debug_info (weights/graph unchanged; the 0.4.0 debug locations broke OS 27 beta 2+ loads) |
ios-ane-h18p/ |
the same ios/ IR, AOT-compiled for the Neural Engine (xcrun coreai-build compile --platform iOS --preferred-compute neural-engine --architecture h18p, coreai-build 3600.83.1 / Xcode 27.0 RC) — 31/31 ANE regions; metadata.json points at the .h18p.aimodelc, tokenizer included; loads through Apple’s unmodified EngineFactory → StaticShapeEngine on iPhone 17-class devices |
557 MB (main-h18p.mlirb sha256 95c85f6c6b8970f9a31dda2b90eac15fee1f3494c316fdfdf7f6b28315811f8d) |
2026-09-15 (conversion/export_ane_stock.py --aot-only) |
ios/ is the portable IR for the ANE bundle (what other repos call ios-static/): compile it yourself for
another chip. The 0.4.0-era IR carried no author/license/description; the AOT step wrote them
(Qwen Team / Apache-2.0) before compiling, so the .aimodelc metadata is complete.
ondevice/_ane_gate)Pending — this session’s device slot comes after S1. The fixture is Qwen/Qwen3-0.6B fp32: alphabet
list (24 steps), no-think chat “What is the capital of France?” (10 steps incl. the stop), 396-id
counting prompt (16 steps). The tool goes RED on a poisoned fixture first. Rules and the transcript
format: ../../knowledge/minicpm5-1b.md §2026-09-15.
Pending (same-day interleaved A/B vs the repo’s macos/ dynamic int4 bundle on the GPU, p128 / g256 n5).
The iOS 27 beta numbers on the HF card (decode 69.6 / 54.1 tok/s, prefill 5,325, warm load 0.045 s,
512p / 1024g, 2026-06) predate the GA toolchain and this AOT; they are not the ANE bundle’s numbers.
# the shipped bundle: AOT of the published IR (download ios/ from the HF repo first)
python3 conversion/zoo_convert.py run qwen3-0.6b-official-ane
# a fresh export of the same recipe with today's toolchain (a different IR, not what is published):
python3 conversion/export_ane_stock.py qwen3-0.6b --devbundle qwen3_0_6b_ane_fresh