qwen3-1.7b-CoreAI-official

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 -->

Qwen3-1.7B — Apple Core AI Export (iPhone GPU)

Pre-converted Apple Core AI (.aimodelc) bundle of Qwen/Qwen3-1.7B, produced with the coreai-models export recipe and presented without modification. Hashes are embedded so the artifact is a reproducible reference point.

This fills the missing 1.7B rung in the dense Qwen3 Core AI line (0.6b / 4b / 8b already exist) — and 1.7B is a meaningful rung: it is the largest dense Qwen3 that still invokes on LiteRT-LM iOS, measured alongside this bundle in a neutral cross-runtime benchmark.

Why GPU-only (no ANE bundle)

The 0.6B repo ships both an ANE (static-shape, palettized) and a GPU (dynamic INT4) bundle. At 1.7B the ANE export is omitted on purpose: the static-shape ANE bundle loads but fails to invoke on iOS 27 — a full benchmark run window produced no output — the same ANE invoke ceiling the 4B static export hits. Rather than ship a bundle that does not run, only the GPU (dynamic INT4) export is published here; it invokes and decodes cleanly on device. (Core AI's GPU path is unaffected — it runs 0.6B/1.7B/4B on iPhone; the ANE path is the one that tops out below 1.7B.)

Bundle

Embedded tokenizer (Qwen/Qwen3-1.7B), 40960 max context. iOS bundles are already AOT-compiled (.aimodelc) for the iPhone 17 Pro GPU target.

Measured — iPhone 17 Pro (iPhone18,1 · iOS 27.0)

Greedy, 128-token budget for short-chat (n=3, iso-cold), 256 for quality. Every figure traces to raw JSONL in the companion benchmark.

The kernel cache persists across launches, so only the first launch after a fresh install is genuinely cold (44.7); once primed it holds ~66 tok/s — the steady state a user actually sees, and the fastest of the runtimes measured at this size (vs MLX Q4 ~66 at 1095 MB, LiteRT-LM int8 ~30 at 512 MB).

Usage

iOS bundles are AOT-compiled and side-loaded into the app container; load via Core AI with the embedded tokenizer and coreai-pipelined GPU engine. See the coreai-models recipe and CoreAIChatMac for an interactive harness.

Provenance

Converted from Qwen/Qwen3-1.7B (Apache-2.0). Quant: dynamic INT4 (linear). Producer: coreai-build-3600.67.5.8.1.

<!-- 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 -->