True2456/Qwen3.8-Flash-Next-ANE-h17

Qwen3.8-Flash-Next ANE graphs (M5 / h17)

Baked Apple Neural Engine MIL + weight blobs for the Rindi Flash-Next serve path. M5 only (h17). Each package is what _ANEInMemoryModel materializes before compileWithQoS:. Another M5 compiles these with the private ANE compiler; an M3/M4 Mac will compile to a different ISA.

Apple's compiled .hwx is a content-addressed local cache and is not a portable file after load. Shipping the bake skips the BF16→int8 quantization (~50 s) and the 360 GB BF16 tree.

These are the attention stack (36 GDN + 12 QSA), not the MoE experts. Pair with True2456/Qwen3.8-Flash-Next-MLX-4bit. Hub id: True2456/Qwen3.8-Flash-Next-ANE-h17.

Bake

Chip	Apple M5 (h17 / H17P)
Decode width	FLASHNEXT_SPEC=4
Prefill width	FLASHNEXT_PREFILL_MIL_K=32 (second procedure, same program)
QSA rungs	256 and 2048

Each package directory is a copy of _ANEInMemoryModel.localModelPath after compile/load:

• model.mil — MIL text
• weights/weight_data.bin — packed int8 projections + fp16 mixers / shared expert
• weights/weight_scale.bin — per-channel scales
• package.json — hexStringIdentifier, sizes

Shared expert bytes are the calibrated 8-bit copy from the MLX 4-bit checkpoint, matching the live server.

Layout

gdn_L{i}/           36 GDN layers (every layer except i%4==3)
qsa_L{i}_front/     12 QSA indexer/mixer fronts
qsa_L{i}_kv256/     decode key-window rung
qsa_L{i}_kv2048/    wide rung; also carries the k=32 prefill procedure
manifest.json