Quantized tencent/Hy3 for Apple Silicon MLX / JANG runtimes — a 295B-total / 21B-active text MoE, packed to ~98 GiB with the native Multi-Token-Prediction (MTP) head preserved. Hy3-JANG2K-MTP is a JANG mixed-affine quantization of Tencent’s Hy3 dense-MoE. It targets Apple Silicon runtimes (MLX / vMLX) and keeps the model’s native MTP layer in the bundle so a compatible engine can run next-token speculative decoding. The 2K profile spends an extra bit on the routed downproj (3-bit vs the 2-bit gate/up), which cleans up the sampling tail relative to a uniform 2-bit pack. Routed-expert effective average: 2.33 bit. AWQ scaling is disabled for this bundle (measured negligible on Hy3). Hy3 is a text-only dense-causal-GQA MoE — not MLA, not SSM, not sliding-window, not a VLM. — 80 decoder layers, hiddensize 4096 — GQA: 64 attention heads / 8 KV heads, headdim 128, QK-norm — RoPE default, ropetheta 11,158,840, maxpositionembeddings 262,144 — MoE: 192 routed experts, top-8, sigmoid router + expert-correction bias, routenorm, routerscalingfactor 2.826, 1 shared expert, firstkdensereplace 1 — MTP: 1 next-token-prediction layer preserved (numnextnpredictlayers = 1) — vocabsize` 120,832 -…
Модальности:
Генерация текста
Области применения:
Логика и рассуждение Диалог / чат Вызов функций (Tool use)
Задача: Генерация текста
Автор: JANGQ-AI
Теги: mlx, hy_v3, hunyuan, hy3, mixture-of-experts, apple-silicon, reasoning, tool-use
Лайков: 4 | Загрузок: 160
Описание основано на материалах HuggingFace. Перевод выполнен автоматически.
