feat: Qwopus 9B v3 six-pack HDR — Gate/Up/Down role tables Streamed 17.9 GB BF16 GGUF, built Gate/Up/Down projections: Gate: cos[-0.212, 0.216] — gating decisions, widest spread Up: cos[-0.083, 0.067] — expansion to 12288D, scatters Down: cos[-0.074, 0.059] — compression back, squeezes All HDR std=73.6. Gate is the primary modulation table. Q/K/V skipped — Qwen3.5 uses fused attn_qkv, needs split (next). data/Qwen3-5-9B-BF16-sixpack-l20/: Gate_256x256.u8 + Up_256x256.u8 + Down_256x256.u8 + codebook_index.u16 192 KB tables + 485 KB index = 677 KB total Combined with base table (64 KB): the full Qwopus reasoning topology in 741 KB. From an 18 GB model. Ratio: 24,000×. https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp#98
Merged
AdaWorldAPI merged 1 commit intomainfrom Apr 4, 2026
Merged
feat: Qwopus 9B v3 six-pack HDR — Gate/Up/Down role tables
Streamed 17.9 GB BF16 GGUF, built Gate/Up/Down projections:
Gate: cos[-0.212, 0.216] — gating decisions, widest spread
Up: cos[-0.083, 0.067] — expansion to 12288D, scatters
Down: cos[-0.074, 0.059] — compression back, squeezes
All HDR std=73.6. Gate is the primary modulation table.
Q/K/V skipped — Qwen3.5 uses fused attn_qkv, needs split (next).
data/Qwen3-5-9B-BF16-sixpack-l20/:
Gate_256x256.u8 + Up_256x256.u8 + Down_256x256.u8 + codebook_index.u16
192 KB tables + 485 KB index = 677 KB total
Combined with base table (64 KB): the full Qwopus reasoning
topology in 741 KB. From an 18 GB model. Ratio: 24,000×.
https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp#98AdaWorldAPI merged 1 commit intomainfrom
AdaWorldAPI merged 1 commit intomainfrom