OsaurusAI/Qwen3.8-27B-JANG_2D

Stamp vmlx_mtp_proposal_head.json (settled 2026-09-03 proposal-head verdict)

Realign safetensors containers for zero-copy mmap loading (tensor bytes unchanged)

card: add measured decode tok/s and best speculation depth

CORRECTION: warm-cache timing — depth 3 is a net loss; best_depth 1/2

CORRECTION: warm-cache timing — depth 3 is a net loss; best_depth 1/2

MTP: validated best_depth=3 (measured tok/s + acceptance curve)

MTP: validated best_depth=3 (measured tok/s + acceptance curve)

stamp measured MTP depth curve (best_depth 1, measured_best_depth 3)

stamp measured MTP depth curve (best_depth 1, measured_best_depth 3)

stamp measured depth-1 MTP acceptance

stamp measured depth-1 MTP acceptance

v2: AWQ + 164k-token calibration + fp16 scale fix + MTP override key fix

model card: drop lineup benchmark table

Qwen3.8-27B JANG_2D — calibrated JANG bundle

model card

model card

initial commit