Core AI model zoo

Qwen3-4B (Apple’s official recipe) — Neural Engine bundle

🤗 mlboydaisuke/qwen3-4b-CoreAI-official · Apache-2.0 · source Qwen/Qwen3-4B

This is not a zoo port: every bundle in the repo is Apple’s own coreai.llm.export qwen3-4b recipe, unmodified (the official/ line — ../../official/README.md, ../../knowledge/apple-models-bench.md). This directory exists because the repo gained a Neural Engine bundle on 2026-09-15 that the zoo gates on the phone; the qwen3-0.6b-official card is the same story one size down.

Bundles

subtree what size provenance
macos/ macOS dynamic, int4 per-block (Apple’s macOS 4bit preset) 2.26 GB 2026-06-11 export, re-uploaded 2026-09-08
ios/ iOS static IR, Apple’s iOS preset for qwen3-4b = models/qwen3/qwen3_4b_mixed_4bit_8bit.yaml (k-means 4-bit, per-grouped-channel group 32 on the body; layers 6/8/11/33/34 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 2.49 GB 2026-06-11 export (coreai-torch 0.4.0), re-serialized 2026-07-21 by strip_debug_info (weights/graph unchanged)
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, 413 s) — 31/31 ANE regions; metadata.json points at the .h18p.aimodelc, tokenizer included 2.66 GB (main-h18p.mlirb 391 MB, sha256 c5dd0b05d8278cd5c42a4f02e60642e2df94d5fbb95c47ccc1325d571aa0a279; delegates 2.1 GB) 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/). The 0.4.0-era IR carried no author/license/description; the AOT step wrote them (Qwen Team / Apache-2.0) before compiling. At 2.66 GB the AOT bundle is above the zoo’s conservative ~1.5 GB iPhone ship line; whether it loads and runs on an iPhone 17 Pro under iOS 27.0 GA is exactly what the device gate below answers (the iOS 27 beta bench loaded it: 13.2 tok/s decode, 3.3 GB footprint, 194 s cold specialization).

Device gate (iPhone 17 Pro, iOS 27.0 24A435, ondevice/_ane_gate)

Pending — this session’s device slot comes after S1. Fixture = Qwen/Qwen3-4B fp32: alphabet list (24 steps, min margin 0.930), no-think chat “What is the capital of France?” (10 steps incl. the stop, min 0.664, EOS 0.999), 396-id counting prompt (16 steps, min 0.958); red twin poisons natural[5]. Rules and transcript format: ../../knowledge/minicpm5-1b.md §2026-09-15.

Measured

Pending (same-day interleaved A/B vs the repo’s macos/ dynamic int4 bundle AOT-compiled for the h18p GPU, p128 / g256 n5). The beta-era HF-card numbers (decode 13.2 / 12.2 tok/s, prefill 546, 512p / 1024g, 2026-06) are not this bundle’s numbers.

Reproduce

python3 conversion/zoo_convert.py run qwen3-4b-official-ane     # AOT of the published ios/ IR (download it first)
python3 conversion/export_ane_stock.py qwen3-4b                  # a fresh export of the same Apple recipe (a different IR)