Card: macOS/iOS 27 GA wording (beta requirement dropped; beta findings dated)

Card: DeviceMark row (2026-09)

Card metadata: library_name coreai, ensure coreai tag

Card: definition line (Core AI, 2026-09)

Card: base_model_relation: quantized (so the conversion lists under the base model's Quantizations, not Finetunes)

qwen3_8_27b_verify_s9_int4lin_d4_hpost: int4 S=9 verify + hidden out for the MTP drafter (29.0 tok/s code, lossless)

qwen3_8_27b_decode_int4lin: int4 text decoder (22.2 tok/s M4 Max, gate 15/16)

qwen3_8_27b_mtp_s9_int8hu_block32_sym: MTP S=9 replay graph (shared-KV sibling)

qwen3_8_27b_mtp_s1_int8hu_block32_sym: MTP head as S=1 stateful drafter (int8)

int4 + lossless MTP ⚡Spec: code 29.0 tok/s, free 22.2 (M4 Max)

Qwen3.8-27B: remove pf32 VL decoder (fp16 overflow on real images; pf16 replaces it)

Qwen3.8-27B: VL decoder pf16 (pf32 chunk overflows fp16 content-dependently)

Link the card back to its collection and the request box

Qwen3.8-27B: qwen3_8_27b_vl_decode_int8hu_block32_sym_pf32

Qwen3.8-27B: qwen3_8_27b_vision_fp16

Qwen3.8-27B: qwen3_8_27b_decode_int8hu_block32_sym

Qwen3.8-27B: config.json

Qwen3.8-27B: LICENSE

Qwen3.8-27B: README.md

initial commit