---
license: other
license_name: wirkung-model-and-runtime-1.0
license_link: LICENSE
base_model:
- Qwen/Qwen3.5-9B
base_model_relation: quantized
library_name: wirkung
pipeline_tag: text-generation
num_parameters: 9410000000
language:
- en
- de
tags:
- wirkung
- origami
- preview
- training-free
- lossless-folding
- constant-memory
- zero-vram
- cpu-inference
- npu
- hcgm
- m-rail
- 9b
- multimodal
- qwen3.5
- hybrid-ssm
---

![Qwen3.5-9B folded without retraining from 18.85 GB to 2.74 GB (6.88x), with a constant 14.82 MB state at any context length.](assets/banner.png)

[Paper](paper/Wirkung_2026_Origami_HCGM.pdf) · [DOI](https://doi.org/10.13140/RG.2.2.31826.72647) · [Runtime (PyPI)](https://pypi.org/project/wirkung/) · [Wirkung](https://wirku.ng) · [Base weights](https://huggingface.co/Qwen/Qwen3.5-9B) · [Contact](mailto:contact@wirku.ng) · [License](LICENSE)

**`Origami-9B` is a 9.41B-parameter Holographic Causal Graph Machine (HCGM) Preview compiled in closed form from `Qwen/Qwen3.5-9B` with zero retraining, zero fine-tuning, and zero gradient steps. It folds the 18.85 GB source checkpoint into 2.74 GB (6.88x 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-9B` preserves the full `Qwen/Qwen3.5-9B` architecture (24 Gated DeltaNet layers, 8 full-attention layers (4 KV groups), 27-layer vision tower, and a 248,320-entry 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`](https://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](paper/Wirkung_2026_Origami_HCGM.pdf).

```bash
pip install -U wirkung && wkng pull wirkung/Origami-9B -o ./Origami-9B
wkng run ./Origami-9B -p "Explain why disk-mapped causal graphs require zero GPU VRAM." -n 128
```

## Source checkpoint against Origami-9B

| Property | Source (`Qwen/Qwen3.5-9B`) | `Origami-9B` (HCGM Preview) |
|:--|--:|--:|
| Parameters | 9.41B | **9.41B** (identical graph topology) |
| Retraining / fine-tuning steps | Pretrained baseline | **0** (closed-form compilation) |
| Checkpoint footprint on disk | 18.85 GB BF16 | **2.74 GB** (**6.88x** smaller; 312 `.o3f` operator 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(<i>T</i>) KV cache (34.40 GB at 1M tokens) | **Constant 14.82 MB O(1) session state at any length** |
| Operator & readout path | Dense per-step matrix scan | Certified Cauchy-Schwarz `lm_head` gate (prunes 99.9996% of 248,320 rows, 0 greedy drift) |
| Runtime dependencies | PyTorch / `vLLM` / `MLX` / CUDA | **`wirkung` + `numpy`** (standalone C11, NPU, `sGPU`, Metal, PTX) |

![Physical byte anatomy and sequence-horizon state scaling of Origami-9B](assets/architecture.png)

Figure 1. (a) Physical byte scale contrasting the 18.85 GB BF16 `Qwen/Qwen3.5-9B` checkpoint pinned in RAM/VRAM against the 2.74 GB `Origami-9B` `.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(<i>T</i>) KV growth against the flat 14.82 MB `Origami-9B` carrier.

## Zero-retraining compilation and constant-state architecture

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

1. **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.
2. **Closed-form spectral folding (`.o3f`, 6.88x without retraining).** Each dense projection is transformed once via a deterministic Rademacher-scrambled 128-point Fast Walsh-Hadamard Transform (`FWHT_128`, generated from SHA-256 -> PCG64 seeds with zero stored sign tables) into a four-class log-bilinear `rANS` stream paired with a rank-<i>r</i> projective crystal (<i>U<sub>r</sub>V<sub>r</sub></i><sup>⊤</sup>). All 312 operators fold from 18.85 GB BF16 to 2.74 GB with zero gradient updates and zero calibration fine-tuning.
3. **Constant 14.82 MB O(1) state synthesized via `Wirkung`.** Derived through `Wirkung`, our internal exact structural-law and closed-form operator synthesis engine operating beyond heuristic symbolic regression, the 24 Gated DeltaNet layers contract via associative WY monoid scans while the 8 full-attention layers replace unbounded Softmax KV caches with a pinned Token-0 sink, an exact 1,024-token local frontier, and key-centered Ward sufficient-statistic centroids. Total live session state stays fixed at 14.82 MB across arbitrary sequence lengths.
4. **Certified readout and online trajectory decomposition (<i>X</i> = <i>X</i><sub>known</sub> + <i>R</i>).** Rather than re-evaluating every parameter blindly for isolated single tokens, HCGM decomposes live hidden trajectories into known low-rank structure <i>X</i><sub>known</sub> (cosine trajectory banks, multi-order suffix tables, and rank-1 Sherman-Morrison-Woodbury updates) plus novel residual <i>R</i>, verifying multiple candidate steps in one weight pass with zero greedy output drift.

![Constant-memory read geometry and folded operator execution in Origami-9B](assets/mechanism.png)

Figure 2. (a) Constant-memory attention read over token coordinate <i>j</i>: pinned sink <i>s</i><sub>0</sub>, bounded Ward + WY Gated DeltaNet far-history carrier, and exact local frontier (<i>W</i> = 1024). (b) Single-pass `.o3f` contraction pairing 4-class Walsh-Hadamard bitplanes with a rank-<i>r</i> projective crystal (<i>U<sub>r</sub>V<sub>r</sub></i><sup>⊤</sup>) and exact Cauchy-Schwarz row pruning.

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

Beyond single-node execution, `Origami-9B` 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 (<i>R</i> ≠ 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.

![Shared M-Rail GPU-less server fleet topology and multi-session memory scaling for Origami-9B](assets/mrail.png)

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 +<i>R</i> solved once commits back to all nodes alongside `HMoE` routing on `cstar`. (b) Total resident RAM/VRAM across <i>N</i> = 1..32 concurrent 32k-token sessions: pinned weights plus <i>N</i> · KV against shared read-only `mmap` plus <i>N</i> · 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.

```bash
pip install -U wirkung
wkng pull wirkung/Origami-9B -o ./Origami-9B
wkng verify ./Origami-9B --full
wkng run ./Origami-9B --chat
wkng serve ./Origami-9B --host 127.0.0.1 --port 8080
```

```python
from wirkung import Engine

eng = Engine("./Origami-9B")
print(eng.generate("Derive the fixed-point condition of a hyperbolic Möbius semigroup:", max_new_tokens=128))
```

## Files

- `layers/*.o3f`, `vocab/*.o3f`: 312 folded HCGM operator files (2.74 GB total bundle)
- `exact_core.safetensors`: exact norms and lightweight structural scalars
- `bundle.json`, `origami9b_manifest.json`: shard inventory and SHA-256 digests verified by `wkng verify`
- `config.json`, `generation_config.json`, `tokenizer.json`, `tokenizer_config.json`, `chat_template.jinja`, `preprocessor_config.json`, `processor_config.json`, `video_preprocessor_config.json`: tokenizer and multimodal processor metadata
- `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](mailto:contact@wirku.ng)
- **X (Twitter):** [https://x.com/FossDT](https://x.com/FossDT)
- **Web:** [https://wirku.ng](https://wirku.ng)

## Papers

- **[The Origami Model Family: Compiling Pretrained Foundation Models into Holographic Causal Graph Machines](paper/Wirkung_2026_Origami_HCGM.pdf)**. David Tom Foss, 2026. DOI [10.13140/RG.2.2.31826.72647](https://doi.org/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](https://doi.org/10.5281/zenodo.19040842)**. David Tom Foss, 2026.
- **[Crystallized Attention: Constant-Memory Exact Store and Exponential-Mode Carriers](https://doi.org/10.5281/zenodo.19058017)**. David Tom Foss, 2026.
- **[Intelligence Is Not Made of Scarcity: Constant-Memory Causal Execution and Associative M-Rail](https://doi.org/10.13140/RG.2.2.28823.43689)**. David Tom Foss, 2026.

## Citation

```bibtex
@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](LICENSE). Compiled in closed form without training steps from Apache-2.0 (`Qwen/Qwen3.5-9B`); full attribution and closed-form transformation notices are recorded in [`NOTICE`](NOTICE).