origami1.1_27b

Weight & constant-state update succeeding wirkung/Origami-27B (v1.0). Compiled with 0 retraining steps from Qwen/Qwen3.8-27B.

Lineage

Component	Source
Predecessor (v1.0)	wirkung/Origami-27B
Base model (64 layers + native MTP)	Qwen/Qwen3.8-27B

Metrics (v1.0 -> v1.1)

Metric	v1.0	v1.1 (origami1.1_27b)
Bundle Size	7.44 GB	11.728 GB
End-to-End ΔNLL vs Base	+0.4268	+0.01147 (37.2x lower)
End-to-End Top-1 Agreement vs Base	76.17%	93.74% (+17.57 pp)
Constant-State ΔNLL vs Exact (1,024 tok)	—	+0.01827 (KL = 0.02009, 96.68% top-1)
Constant-State Passkey Recall (T=640, T=2048)	—	2/3 (3/3 exact mode)
Live KV Allocation Growth (T=1K..128K)	0.0 B/tok	0.0 B/tok (ΔRSS = 0.0 MB)
CPU Decode (x86_64 AVX2, 16 threads, constant-state)	—	4.74 tok/s (15.10 tok/s prefill · 6.97 tok/s MTP)

Contents

• layers/layer-000..layers/layer-063: 400 .o3f operator weights
• vocab/: vocabulary & projection weights
• mtp/: speculative MTP module
• assets/o1_crystal_v2_27b.bin: constant-memory state sidecar
• exact_core.safetensors, bundle.json, config.json, tokenizer files, LICENSE, NOTICE

Download

huggingface-cli download wirkung/origami1.1_27b --local-dir ./origami1.1_27b