Qwen3.8-27B-CoreAI

Core AI is Apple's on-device ML runtime in iOS 27 / macOS 27 and the successor to Core ML: PyTorch models are exported with Apple's coreai-torch (LLMs: coreai.llm.export) into .aimodel bundles that run on the GPU or the Neural Engine, e.g. Qwen3-8B 4-bit decodes at 94 tok/s on an M4 Max GPU, MLX 90 under the same protocol (apple-silicon-llm-bench, macOS 27 beta 26A5353q, 2026-06-11).

<!-- gen-cards:devicemark begin (managed by scripts/gen-cards + tools/devicemark_row.py — edit cards.json, not this block) -->

This model has no row on DeviceMark, the on-device LLM leaderboard.

<!-- gen-cards:devicemark end -->

Qwen3.8-27B — Apple Core AI (.aimodel)

The Qwen3.8 generation's dense 27B, converted to Apple's Core AI (the Core ML successor announced at WWDC26) — ported the day the weights landed. This repo ships the full VLM (text decoder + vision path), an int4 text decoder, and the checkpoint's own MTP head as a lossless ⚡Spec drafter pair. The text decoder is the Qwen3.5 hybrid graph run dense, 64 layers on a 3:1 interleave of GatedDeltaNet linear-attention mixers (GVA 48v/16k) and gated full attention (24 q / 4 KV, head_dim 256), untied 248 320-vocab head, 262 K native context. It rides Apple's coreai-pipelined GPU engine decode-only and loop-free, with the SSM conv/recurrent states carried as fixed-shape extra states. The vision path adds the 458M ViT tower and an embeddings-input decoder variant with real interleaved mRoPE (see below).

This is a reasoning model — the chat template opens a <think> span and generations spend their first tokens thinking. Budget max-tokens accordingly.

Mac-class, Mac-only: 18–28 GB is far past the iPhone memory ceiling. On an M4 Max the whole 27B is read per token — memory-bandwidth-bound by construction.

Requires macOS 27 (Core AI ships with the OS). Conversion code, gates and knowledge base: coreai-model-zoo.

Bundles

int8 text row: M4 Max 128 GB, macOS 27 beta, release llm-benchmark -p 64 -g 128 -n 3, COREAI_CHUNK_THRESHOLD=1. int4 and ⚡Spec rows: same machine, measured in the zoo's CoreAIChatMac engine (cool state, GPU solo, medians over the generation; the engine reproduces the int8 llm-benchmark number at 15.6). Eager quant gates: teacher-forced single-step argmax vs the HF bf16 oracle under the margin ≥ 0.1 rule — int8 15/16, int4 15/16, both missing only the same 0.061-margin knife-edge; the fp16 full-precision control is 16/16. Engine transcript in the zoo card directory.

Vision rows: same machine, python runtime on the AOT h16c compile (command below). Prefill is 5× the text bundle's because the VLM decoder is a _pf32 multifunction bundle — a static S=32 "prefill" function chunks the prompt while "main" (S=1) decodes; image prompts are ~316 tokens, so this is what makes the image path usable. Suite gate: 6 cases (3 COCO images × 2 coarse prompts, one text-before-image) against the bf16 HF oracle, greedy 24 tokens, full-chain (NumPy preprocess → tower → embed splice → decoder). The fp16 eager control on the mixed text+image sequences is 32/32 token-exact.

No iPhone numbers are published here because none were measured (18 GB is still far past the iPhone ceiling; the tower alone would fit but has no on-device decoder to feed).

⚡Spec — lossless speculative decoding with the checkpoint's own MTP head

The checkpoint ships a trained 1-layer MTP draft head (mtp.*, DeepSeek-V3 style: fc([emb|hidden]) → decoder layer → norm, shared lm_head). This repo includes it as a Core AI drafter pair. The loop is lossless by construction — draft tokens are verified by the target graph and the committed text is byte-identical to plain greedy in every measured run.

Three bundles cooperate:

• verify_s9_int4lin_d4_hpost — static S=9 verify forward: 1 committed + 8 draft tokens in, per-position logits plus the last-layer hidden out (the MTP head's input). Chunk-scan doublings cut to 4 (exact for S=9).
• mtp_s1_... — the MTP head as an S=1 stateful drafter with its own 1-layer KV.
• mtp_s9_... — the same weights as an S=9 graph sharing that KV, so the host replays committed context into the drafter in one forward instead of nine (worth +33% on code).

Measured end-to-end (M4 Max, 256-token cap, cool state, GPU solo):

Spec pays where the head predicts well (code: 4.7 accepted/round) and loses on free-form (1.5/round): the GDN verify forward carries a fixed ~60 ms chunk-scan cost that quantization does not shrink (109 ms verify vs 45 ms decode step). So the ship shape is split: verify+MTP pair for code-like work, plain decode bundle for prose. The int8 pair measures 26.7 tok/s on code (1.71× its 15.6 decode) — int4 wins both absolute and split. Reference host loop (replay → seed → K draft steps → verify → commit, snapshot/re-anchor state discipline): the zoo repo's CoreAIChatMac SpecDecodeEngine. The verify bundle's metadata names its drafter pair (spec_draft, spec_mtp_replay) for auto-pairing.

The vision path, in one paragraph

The tower is a fixed-grid one-shot encoder: patches [1024, 1536] → image_embeds [256, 5120] at a baked 512×512 tile (32×32 patches, 2×2 merge — the fixed square grid stretches non-square images). The host resizes/normalizes/patchifies in NumPy (_smoke/qwen38vl_preprocess.py, gated exactly against the HF processor), runs the tower once per image, gathers text-token rows from the shipped embed_tokens.safetensors (2.5 GB, fp16), splices tower rows at the 256 <|image_pad|> positions, and feeds the result to the decoder's inputs_embeds input together with three int32 mRoPE position planes (pos_t/pos_h/pos_w — text ramps, image tokens self-locate on the merged grid, an image consumes only max(H,W)/2 = 16 rope positions; _smoke/qwen38vl_host.py is the reference host, asserted against the oracle's captured positions). Text-only prompts make the three planes equal and the graph reduces to plain partial RoPE — i.e. the same numerics as the text bundle.

llm-runner/llm-benchmark cannot drive this bundle (embeddings and rope planes are not engine inputs); the reference driver is _smoke/test_qwen38vl_suite_gate.py. Driving it from the python runtime needs the AOT compile (the JIT path asserts in MPSGraph's ANE region pass on this multifunction graph):

xcrun coreai-build compile qwen3_8_27b_vl_decode_int8hu_block32_sym_pf32.aimodel \
    --platform macOS --preferred-compute gpu --expect-frequent-reshapes --architecture h16c

Run it

git clone https://github.com/john-rocky/coreai-kit
cd coreai-kit/Examples/ChatDemo
swift run chat-cli --model qwen3.8-27b --prompt "What can you do, offline?"

Or in Swift, via CoreAIKit:

import CoreAIKit
let chat = try await ChatSession(catalog: "qwen3.8-27b")
let reply = try await chat.respond(to: prompt)

Reproduce

git clone https://github.com/john-rocky/coreai-model-zoo
cd coreai-model-zoo
python3 conversion/zoo_convert.py run qwen3.8-27b

Recipes (text): export_qwen3_5_decode_pipelined.py int8hu --head-sym --hf-id Qwen/Qwen3.8-27B (int8 ship recipe, identical to Qwen3.6-27B) and ... int4lin --hf-id Qwen/Qwen3.8-27B (int4). ⚡Spec set: export_qwen3_5_verify_pipelined.py int4lin --hf-id Qwen/Qwen3.8-27B --s 9 --doublings 4 --emit-hidden post (verify) + export_qwen3_8_mtp.py (S=1 drafter; --s 9 for the replay sibling). Recipe (vision path): export_qwen38vl_pipelined.py int8hu — one run emits the fp16 tower AND the pf32 VLM decoder (+ embed_tokens.safetensors). Port write-up: knowledge/qwen3.8-27b-port.md.

<!-- funnel:v1 -->

---

More models in this format: Core AI Model Zoo — 75 models, each with the recipe that produced it.

Want a different model on-device? Open a request — free, open weights only; the export and its measured numbers get published publicly.

<!-- /funnel:v1 -->