origami1.1_9b

Weight & constant-state update succeeding wirkung/Origami-9B (v1.0). Compiled with 0 retraining steps (text + vocabulary + MTP).

Lineage

Component	Source
Predecessor (`v1.0`)	`wirkung/Origami-9B`
Backbone & vocabulary (`32` layers)	`xtools-at/DeepSeek-V4-Pro-Qwen3.5-9B-UNCENSORED-heretic`
Speculative MTP head (`mtp/`)	`Qwen/Qwen3.5-9B`

Metric	`v1.0`	`v1.1` (`origami1.1_9b`)
Bundle Size (Text + Vocab + MTP)	`2.74 GB`	`4.123 GB`
End-to-End `ΔNLL` vs Base	`+0.2044`	`+0.0294` (`6.95x` lower)
End-to-End Top-1 Agreement vs Base	`76.89%`	`85.73%` (`+8.84 pp`)
Constant-State `ΔNLL` vs Exact (`1,024` tok)	`—`	`+0.00564` (`KL = 0.02364`, `94.13%` top-1)
Constant-State Passkey Recall (`T=640`, `T=2048`)	`—`	`3/3` (`100%`)
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)	`—`	`14.58 tok/s` (`47.39 tok/s` prefill · `17.21 tok/s` MTP)

layers/layer-000..layers/layer-031: 200 .o3f operator weights
vocab/: vocabulary & projection weights
mtp/: speculative MTP module
assets/o1_crystal_v2_9b.bin: constant-memory state sidecar
exact_core.safetensors, bundle.json, config.json, tokenizer files, LICENSE, NOTICE

huggingface-cli download wirkung/origami1.1_9b --local-dir ./origami1.1_9b

Base model

Finetuned

Quantized

(725)

this model