wirkung/Origami-31B · Hugging Face

Paper · DOI · Runtime (PyPI) · Wirkung · Base weights · Contact · License

Origami-31B is a 32.68B-parameter Holographic Causal Graph Machine (HCGM) Preview compiled in closed form from google/gemma-4-31b-it with zero retraining, zero fine-tuning, and zero gradient steps. It folds the 62.55 GB source checkpoint into 11.56 GB (5.41x compression), executes directly from NVMe/SSD storage via OS mmap with 0 GB GPU VRAM and 0 GB dedicated weight-RAM, and keeps live state strictly constant (O(1)) across arbitrary horizons on CPU, NPU, sGPU, Metal, and standalone PTX.

Origami-31B preserves the full google/gemma-4-31b-it architecture (50 sliding-window attention layers, 10 symmetric K = V global-attention layers, a 27-layer SigLIP vision tower, and a tied 262,144-entry softcapped vocabulary) while turning a passive weight checkpoint into a self-contained binary file executed under one cross-silicon contract. It succeeds our first-generation constant-state releases published under huggingface.co/tfwnotops by removing vLLM, MLX, and Transformers runtime dependencies altogether. This repository is a Preview release: structural byte footprints, fold ratios, and constant-memory invariants below are fixed properties of the compiled artifact, while hardware-dependent throughput figures are omitted from the card and documented in the Origami Family Paper.

pip install -U wirkung && wkng pull wirkung/Origami-31B -o ./Origami-31B
wkng run ./Origami-31B -p "Explain why monotone logit softcapping preserves exact Cauchy-Schwarz row pruning." -n 128

Source checkpoint against Origami-31B

Property	Source (`google/gemma-4-31b-it`)	`Origami-31B` (HCGM Preview)
Parameters	32.68B	32.68B (identical graph topology)
Retraining / fine-tuning steps	Pretrained baseline	0 (closed-form compilation)
Checkpoint footprint on disk	62.55 GB BF16	11.56 GB (5.41x smaller; 601 `.o3f` text, vision, and vocabulary files + `exact_core.safetensors`)
Required GPU VRAM	Multi-tens of GB HBM	0 GB (runs on plain CPU / NPU / `sGPU` or optional GPU)
Dedicated weight-RAM	Pinned in host / device RAM	0 GB (read-only OS `mmap` directly from SSD/NVMe)
Live session / horizon state	Linearly growing O(T) global KV cache across 10 full-attention layers	Constant O(1) bounded sliding-window + symmetric K = V carrier at any length
Operator & readout path	Dense per-step matrix scan	Certified Cauchy-Schwarz gate under κ tanh(z/κ) logit softcapping (median 2.0 of 262,144 rows, 0 drift)
Runtime dependencies	PyTorch / `vLLM` / `MLX` / CUDA	`wirkung` + `numpy` (standalone C11, NPU, `sGPU`, Metal, PTX)

Figure 1. (a) Physical byte scale contrasting the 62.55 GB BF16 google/gemma-4-31b-it checkpoint against the 11.56 GB Origami-31B .o3f bundle mapped directly from SSD (0 GB GPU VRAM, 0 GB dedicated weight-RAM). (b) Live conversation state across 4k to 1M tokens: linear O(T) KV growth against the bounded O(1) Origami-31B carrier.

Zero-retraining compilation and constant-state architecture

Every Origami model is produced by the same four-part compilation pipeline without gradient descent:

The file is the machine (disk-resident mmap execution). An HCGM model is an immutable binary artifact on disk. The operating system maps operator pages read-only on demand from SSD or NVMe storage, requiring neither GPU VRAM nor weight-deserialization buffers in host RAM. Under our v3-exakt dyadic integer reduction contract, CPU (AVX2 / NEON), general NPUs, embedded sGPU (cpu_cuda), Apple Metal, and driver-API PTX evaluate identical canonical reductions rather than drifting floating-point kernel approximations.
Closed-form multimodal folding (.o3f, 5.41x without retraining). Compiled directly from the pure BF16 62.55 GB google/gemma-4-31b-it repository (61.40 GB active text and vision weights) with zero gradient steps, Origami-31B packs all 60 decoder layers, the 27-layer SigLIP vision encoder, and the tied 262,144-row head into 11.56 GB across 601 .o3f operator files via Rademacher-scrambled FWHT_128 bitplanes and rank-r projective crystals (U_rV_r^⊤).
Constant O(1) symmetric K = V carrier synthesized via Wirkung. Through Wirkung, our internal exact structural-law and closed-form operator synthesis engine, the 50 local sliding-window layers pair with a constant-capacity pinned-sink carrier on the 10 global symmetric K = V attention layers, holding live memory strictly independent of sequence length T while preserving exact argmax ordering across monotonically softcapped ψ(z) = 30 tanh(z/30) vocabulary logits.
Certified readout and online trajectory decomposition (X = X_known + R). Rather than re-evaluating every parameter blindly for isolated single tokens, HCGM decomposes live hidden trajectories into known low-rank structure X_known (cosine trajectory banks, multi-order suffix tables, and rank-1 Sherman-Morrison-Woodbury updates) plus novel residual R, verifying multiple candidate steps in one weight pass with zero greedy output drift.

Figure 2. (a) Constant-memory attention read over token coordinate j: pinned sink s₀, bounded symmetric K = V far-history carrier, and exact sliding-window frontier. (b) Single-pass .o3f contraction with 4-class Walsh-Hadamard bitplanes, rank-r projective crystal (U_rV_r^⊤), and softcap-invariant Cauchy-Schwarz pruning.

Shared trajectory assimilation (M-Rail), GPU-less server fleets and multi-model routing (HMoE)

Beyond single-node execution, Origami-31B attaches natively to Shared M-Rail (.mrail v3), a content-addressed associative trajectory manifold shared across processes or across an enterprise network:

Warm up once on one GPU, run permanently on disk-only CPU/NPU servers. Because HCGM models require only SSD storage and megabytes of live state, an organization can run a single GPU (or a short temporary cloud GPU lease) until its recurring domain workflows are crystallized into a shared .mrail file, then shut the GPU down and serve frontier models across commodity CPU/NPU VMs and CPU-only servers. Known trajectories execute via Track copy morphisms or resume in O(1) space from Station (ANCHOR + DELTA) snapshots; whenever any server encounters an unseen task (R ≠ 0), computing and committing that residual once makes the solution immediately available to every other node on the rail.
Heavy Mixture of Experts (HMoE). Because idle disk-mapped models consume zero weight-RAM, M-Rail routes across complete compiled Origami models (Origami-9B, Origami-27B, Origami-31B, Origami-78B, Origami-H3), zero-recompile static-graph NPU specialist pools, and sub-2 KB exact algebraic group organs on a synchronous station bus (cstar) under deterministic check gates.

Figure 3. (a) Shared M-Rail (.mrail v3) fleet architecture: optional temporary GPU warmup feeds a permanent zero-GPU CPU/NPU/sGPU server fleet where any novel residual +R solved once commits back to all nodes alongside HMoE routing on cstar. (b) Total resident RAM/VRAM across N = 1..32 concurrent 32k-token sessions: pinned weights plus N · KV against shared read-only mmap plus N · O(1) constant state.

How to run

Requires Python 3.10+ and the wirkung runtime package (pip install -U wirkung), which installs the wkng, wirkung, sgpu, and rail CLI entrypoints.

pip install -U wirkung
wkng pull wirkung/Origami-31B -o ./Origami-31B
wkng verify ./Origami-31B --full
wkng run ./Origami-31B --chat
wkng serve ./Origami-31B --host 127.0.0.1 --port 8080

from wirkung import Engine

eng = Engine("./Origami-31B")
print(eng.generate("Compare linear O(T) KV caches against fixed-capacity symmetric K=V carriers:", max_new_tokens=128))

Files

layers/*.o3f, vision/*.o3f, vocab/*.o3f: 601 folded HCGM text, vision, and vocabulary operator files (11.56 GB total bundle)
exact_core.safetensors: exact norms and structural scalars
bundle.json: shard inventory and SHA-256 digests verified by wkng verify
config.json, generation_config.json, processor_config.json, tokenizer.json, tokenizer_config.json, chat_template.jinja: multimodal and tokenizer configuration
paper/Wirkung_2026_Origami_HCGM.pdf: the Origami Model Family technical report
assets/: banner.png, architecture.png, mechanism.png, mrail.png
LICENSE, NOTICE: Wirkung Model and Runtime License 1.0 and upstream attribution

Inquiries and collaboration

We welcome inquiries regarding custom zero-retraining HCGM model compilation, closed-form formula and exact algorithm synthesis via Wirkung, Shared M-Rail / zero-GPU enterprise fleet deployments, and research collaboration:

Email: contact@wirku.ng
X (Twitter): https://x.com/FossDT
Web: https://wirku.ng

Papers

The Origami Model Family: Compiling Pretrained Foundation Models into Holographic Causal Graph Machines. David Tom Foss, 2026. DOI 10.13140/RG.2.2.31826.72647. Included directly in this repository at paper/Wirkung_2026_Origami_HCGM.pdf.
Holographic Causal Graph Machines: Source-Free Constant-State Execution of Pretrained Transformers. David Tom Foss, 2026.
Crystallized Attention: Constant-Memory Exact Store and Exponential-Mode Carriers. David Tom Foss, 2026.
Intelligence Is Not Made of Scarcity: Constant-Memory Causal Execution and Associative M-Rail. David Tom Foss, 2026.

Citation

@techreport{foss2026origami,
  title       = {The Origami Model Family: Compiling Pretrained Foundation Models into Holographic Causal Graph Machines},
  author      = {Foss, David Tom},
  institution = {Wirkung Inc.},
  year        = {2026},
  month       = nov,
  doi         = {10.13140/RG.2.2.31826.72647},
  url         = {https://doi.org/10.13140/RG.2.2.31826.72647}
}

License

Released under the Wirkung Model and Runtime License 1.0. Compiled in closed form without training steps from Gemma Terms of Use / upstream license (google/gemma-4-31b-it); full attribution and closed-form transformation notices are recorded in NOTICE.

Downloads last month: -