wolfram-pro/qwen3.5-4b-coreai

Qwen3.5-4B — Core AI export for iPhone and iPad

Own conversion of Qwen/Qwen3.5-4B (revision 851bf6e806efd8d0a36b00ddf55e13ccb7b8cd0a, Apache-2.0) to Apple Core AI (iOS/macOS 27): GDN hybrid graph, static input_ids [1, 1], context 4096, int4 per-block-32 body with GPTQ, int8 down_proj, shared int8 embedding/head table.

• q4_gptq_b32_down8/ — portable .aimodel bundle (3.09 GB), specialized by the host.
• <arch>/ — closures compiled ahead of time for one iOS GPU architecture (h13g M1, h14g M2, h15g M3, h16g M4, h16p A17 Pro, h17g M5, h17p A18 / A18 Pro, h18p A19 / A19 Pro, h19p A20 Pro), iOS 27.2+. Each folder carries a descriptor.json with the size and SHA-256 of every file.

Tokenizer files are the upstream ones, unchanged. The LICENSE file is the upstream Apache-2.0 license. Reproduction scripts and measurements live with the consuming application; nothing here is an official Qwen release.