Update unified Origami Preview model card, wirkung.cls figures, and Family Paper PDF
Browse files- .gitattributes +4 -0
- README.md +81 -83
- assets/{weights_vs_origami.png → architecture.png} +2 -2
- assets/architecture_organs.pdf +0 -0
- assets/banner.png +2 -2
- assets/{architecture_organs.png → mechanism.png} +2 -2
- assets/{mrail_network.png → mrail.png} +2 -2
- assets/mrail_network.pdf +0 -0
- assets/weights_vs_origami.pdf +0 -0
- paper/Foss_2026_Origami_HCGM.pdf +3 -0
.gitattributes
CHANGED
|
@@ -345,3 +345,7 @@ assets/banner.png filter=lfs diff=lfs merge=lfs -text
|
|
| 345 |
assets/keyframes.png filter=lfs diff=lfs merge=lfs -text
|
| 346 |
assets/mrail_network.png filter=lfs diff=lfs merge=lfs -text
|
| 347 |
assets/weights_vs_origami.png filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 345 |
assets/keyframes.png filter=lfs diff=lfs merge=lfs -text
|
| 346 |
assets/mrail_network.png filter=lfs diff=lfs merge=lfs -text
|
| 347 |
assets/weights_vs_origami.png filter=lfs diff=lfs merge=lfs -text
|
| 348 |
+
paper/Foss_2026_Origami_HCGM.pdf filter=lfs diff=lfs merge=lfs -text
|
| 349 |
+
assets/architecture.png filter=lfs diff=lfs merge=lfs -text
|
| 350 |
+
assets/mechanism.png filter=lfs diff=lfs merge=lfs -text
|
| 351 |
+
assets/mrail.png filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -4,140 +4,138 @@ license_name: wirkung-model-and-runtime-1.0
|
|
| 4 |
license_link: LICENSE
|
| 5 |
library_name: wirkung
|
| 6 |
pipeline_tag: text-to-video
|
| 7 |
-
|
| 8 |
-
- MiniMaxAI/MiniMax-H3
|
| 9 |
-
base_model_relation: quantized
|
| 10 |
language:
|
| 11 |
- en
|
| 12 |
-
-
|
| 13 |
tags:
|
| 14 |
- wirkung
|
| 15 |
- origami
|
| 16 |
- preview
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 17 |
- hcgm
|
| 18 |
- m-rail
|
| 19 |
-
-
|
| 20 |
-
-
|
| 21 |
-
- constant-state
|
| 22 |
-
- text-to-video
|
| 23 |
- audio-video
|
| 24 |
-
-
|
| 25 |
-
- zero-gpu
|
| 26 |
---
|
| 27 |
|
| 28 |
-
`) across continuous uncut horizons.</sub>
|
| 53 |
|
| 54 |
-
|
| 55 |
|
| 56 |
-
|
| 57 |
-
- **Parseval Spectral Stabilization on Nonlinear Flow Blocks**: Naive low-bit compression collapses coupled audio-video flow trajectories inside nonlinear SwiGLU blocks. Origami-H3 enforces Parseval row-norm conservation paired with projective spectral crystals so flow velocities remain coherent from pure Gaussian noise to decoded media.
|
| 58 |
-
- **Continuous Uncut Horizon in Flat `O(1)` State**: Rather than resetting latent state at short clip boundaries, `M-Rail` sideline observers and smooth boundary continuation maintain temporal motion and acoustic phase continuity inside a constant memory envelope.
|
| 59 |
|
| 60 |
-
|
| 61 |
-
<sub><b>Figure 2.</b> Three-stage full-width HCGM topology of Origami-H3: Parseval-stabilized conditioning and folded hypernetwork gates, coupled audio-video flow `DiT` with in-graph `Turbo-LoRA`, and streaming spatiotemporal `VAE` readout for continuous uncut takes.</sub>
|
| 62 |
|
| 63 |
-
|
|
|
|
|
|
|
|
|
|
| 64 |
|
| 65 |
-

|
| 120 |
|
| 121 |
## Papers
|
| 122 |
|
| 123 |
-
- **[The
|
| 124 |
-
- **[
|
| 125 |
-
- **[Crystallized Attention:
|
| 126 |
-
- **[
|
| 127 |
|
| 128 |
## Citation
|
| 129 |
|
| 130 |
```bibtex
|
| 131 |
-
@
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
|
| 135 |
-
year
|
| 136 |
-
|
| 137 |
-
|
| 138 |
}
|
| 139 |
```
|
| 140 |
|
| 141 |
## License
|
| 142 |
|
| 143 |
-
|
|
|
|
| 4 |
license_link: LICENSE
|
| 5 |
library_name: wirkung
|
| 6 |
pipeline_tag: text-to-video
|
| 7 |
+
num_parameters: 21600000000
|
|
|
|
|
|
|
| 8 |
language:
|
| 9 |
- en
|
| 10 |
+
- de
|
| 11 |
tags:
|
| 12 |
- wirkung
|
| 13 |
- origami
|
| 14 |
- preview
|
| 15 |
+
- training-free
|
| 16 |
+
- lossless-folding
|
| 17 |
+
- constant-memory
|
| 18 |
+
- zero-vram
|
| 19 |
+
- cpu-inference
|
| 20 |
+
- npu
|
| 21 |
- hcgm
|
| 22 |
- m-rail
|
| 23 |
+
- 22b
|
| 24 |
+
- video-generation
|
|
|
|
|
|
|
| 25 |
- audio-video
|
| 26 |
+
- diffusion-transformer
|
|
|
|
| 27 |
---
|
| 28 |
|
| 29 |
+

|
| 30 |
|
| 31 |
+
[Paper](paper/Foss_2026_Origami_HCGM.pdf) · [Runtime (PyPI)](https://pypi.org/project/wirkung/) · [Wirkung](https://wirku.ng) · [Base weights](https://github.com/MiniMax-AI) · [Contact](mailto:contact@wirku.ng) · [License](LICENSE)
|
| 32 |
|
| 33 |
+
**`Origami-H3` is a 21.60B-parameter Holographic Causal Graph Machine (HCGM) Preview compiled in closed form from `MiniMax-H3` with zero retraining, zero fine-tuning, and zero gradient steps. It folds the 44.89 GB source checkpoint into 10.38 GB (4.33x 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.**
|
| 34 |
+
|
| 35 |
+
`Origami-H3` preserves the full `MiniMax-H3` architecture (50 joint audio-video Diffusion Transformer (DiT) blocks, 2 token refiners, a fused rank-128 Walsh-6 Turbo-LoRA, a 24-block 3D `SoL-H3-VAE`, and an `Audio-VAE`) 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/Foss_2026_Origami_HCGM.pdf).
|
| 36 |
|
| 37 |
```bash
|
| 38 |
pip install -U wirkung && wkng pull wirkung/Origami-H3 -o ./Origami-H3
|
| 39 |
+
wkng verify ./Origami-H3 --full && wkng info ./Origami-H3
|
| 40 |
```
|
| 41 |
|
| 42 |
+
## Source checkpoint against Origami-H3
|
| 43 |
|
| 44 |
+
| Property | Source (`MiniMax-H3`) | `Origami-H3` (HCGM Preview) |
|
| 45 |
+
|:--|--:|--:|
|
| 46 |
+
| Parameters | 21.60B | **21.60B** (identical graph topology) |
|
| 47 |
+
| Retraining / fine-tuning steps | Pretrained baseline | **0** (closed-form compilation) |
|
| 48 |
+
| Checkpoint footprint on disk | 44.89 GB BF16 | **10.38 GB** (**4.33x** smaller; 307 `.o3f` DiT + 3D/Audio VAE operator files + `exact_core.safetensors`) |
|
| 49 |
+
| Required GPU VRAM | Multi-tens of GB HBM | **0 GB** (runs on plain CPU / NPU / `sGPU` or optional GPU) |
|
| 50 |
+
| Dedicated weight-RAM | Pinned in host / device RAM | **0 GB** (read-only OS `mmap` directly from SSD/NVMe) |
|
| 51 |
+
| Live session / horizon state | Unbounded multi-window 3D spatiotemporal latent buffers (exceeds 16 GB unified memory) | **Constant $O(1)$ horizon carrier across continuous uncut audio-video windows** |
|
| 52 |
+
| Operator & readout path | Dense per-step matrix scan | 4-step shifted Euler flow matching + `M-Rail` physical/geometric sideline verification |
|
| 53 |
+
| Runtime dependencies | PyTorch / `vLLM` / `MLX` / CUDA | **`wirkung` + `numpy`** (standalone C11, NPU, `sGPU`, Metal, PTX) |
|
| 54 |
|
| 55 |
+

|
|
|
|
| 56 |
|
| 57 |
+
Figure 1. (a) Physical byte scale contrasting the 44.89 GB BF16 `MiniMax-H3` pipeline against the 10.38 GB `Origami-H3` `.o3f` bundle mapped directly from SSD (`0 GB` dedicated GPU VRAM, `0 GB` weight-RAM). (b) Live spatiotemporal state across consecutive latent audio-video windows: unbounded window caching against the constant $O(1)$ `Origami-H3` horizon state.
|
| 58 |
|
| 59 |
+
## Zero-retraining compilation and constant-state architecture
|
|
|
|
|
|
|
| 60 |
|
| 61 |
+
Every Origami model is produced by the same four-part compilation pipeline without gradient descent:
|
|
|
|
| 62 |
|
| 63 |
+
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.
|
| 64 |
+
2. **Closed-form Parseval + projective-crystal folding (`.o3f`, 4.33x without retraining).** Compiled directly from the frozen 44.89 GB BF16 `MiniMax-H3` pipeline with zero gradient steps into 10.38 GB across 307 `.o3f` operator files, `Origami-H3` pairs Parseval row-norm conservation with rank-48 boundary crystals (blocks 0..2 and 47..49), rank-32 interior crystals (blocks 3..46), gauge-balanced QR-SVD rank-128 Walsh-6 `rANS` Turbo-LoRA, and a folded 97-operator 3D `SoL-H3-VAE`, preserving nonlinear SwiGLU velocity fields on 16 GB consumer hardware.
|
| 65 |
+
3. **Constant $O(1)$ uncut horizon state synthesized via `Wirkung`.** Using `Wirkung`, our internal exact structural-law and closed-form operator synthesis engine, `Origami-H3` replaces hard scene-cut window resets with a bounded $O(1)$ spatiotemporal carrier across consecutive 4-step shifted Euler windows while parallel constant-state `M-Rail` observers check 3D geometric, physical, and audio-visual invariants in the same weight stream.
|
| 66 |
+
4. **Certified readout and online trajectory decomposition ($X = X_{\mathrm{known}} + R$).** Rather than re-evaluating every parameter blindly for isolated single tokens, HCGM decomposes live hidden trajectories into known low-rank structure $X_{\mathrm{known}}$ (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.
|
| 67 |
|
| 68 |
+

|
|
|
|
| 69 |
|
| 70 |
+
Figure 2. (a) Continuous multi-window trajectory coherence under uncut `Origami-H3` `M-Rail` flow against open-loop window-boundary drift. (b) Single-pass `.o3f` DiT operator combining Parseval `FWHT_128` 4-class bitplanes and Walsh-6 LoRA with a rank-$r$ projective crystal ($U_r V_r^\top$) feeding the folded 3D and Audio VAEs.
|
| 71 |
|
| 72 |
+
## Shared trajectory assimilation (M-Rail), GPU-less server fleets and multi-model routing (HMoE)
|
|
|
|
|
|
|
| 73 |
|
| 74 |
+
Beyond single-node execution, `Origami-H3` attaches natively to **Shared `M-Rail`** (`.mrail` v3), a content-addressed associative trajectory manifold shared across processes or across an enterprise network:
|
| 75 |
|
| 76 |
+
- **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 `Spur` copy morphisms or resume in $O(1)$ space from `Station` (`ANCHOR` + `DELTA`) snapshots; whenever any server encounters an unseen task ($R \neq 0$), computing and committing that residual once makes the solution immediately available to every other node on the rail.
|
| 77 |
+
- **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.
|
| 78 |
|
| 79 |
+

|
| 80 |
+
|
| 81 |
+
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 $N \cdot \mathrm{KV}$ against shared read-only `mmap` plus $N \cdot O(1)$ constant state.
|
| 82 |
+
|
| 83 |
+
## How to run
|
| 84 |
|
| 85 |
+
Requires Python 3.10+ and the `wirkung` runtime package (`pip install -U wirkung`), which installs the `wkng`, `wirkung`, `sgpu`, and `rail` CLI entrypoints.
|
| 86 |
+
|
| 87 |
+
```bash
|
| 88 |
+
pip install -U wirkung
|
| 89 |
+
wkng pull wirkung/Origami-H3 -o ./Origami-H3
|
| 90 |
+
wkng info ./Origami-H3
|
| 91 |
+
wkng verify ./Origami-H3 --full
|
| 92 |
```
|
| 93 |
|
| 94 |
```python
|
| 95 |
from wirkung import Engine
|
| 96 |
|
| 97 |
+
eng = Engine("./Origami-H3")
|
| 98 |
+
print(eng.info())
|
|
|
|
|
|
|
|
|
|
|
|
|
| 99 |
```
|
| 100 |
|
| 101 |
+
## Files
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 102 |
|
| 103 |
+
- `layers/*.o3f`, `vae/*.o3f`: 307 folded HCGM DiT, Walsh-6 LoRA, 3D `SoL-H3-VAE`, and `Audio-VAE` operator files (10.38 GB total bundle)
|
| 104 |
+
- `exact_core.safetensors`: exact modulation, norm, and conditioning scalars
|
| 105 |
+
- `bundle.json`: shard inventory and SHA-256 digests verified by `wkng verify`
|
| 106 |
+
- `config.json`, `generation_config.json`: pipeline and 4-step Euler schedule configuration
|
| 107 |
+
- `paper/Foss_2026_Origami_HCGM.pdf`: the Origami Model Family technical report
|
| 108 |
+
- `assets/`: `banner.png`, `architecture.png`, `mechanism.png`, `mrail.png`
|
| 109 |
+
- `LICENSE`, `NOTICE`: Wirkung Model and Runtime License 1.0 and upstream attribution
|
| 110 |
|
| 111 |
+
## Inquiries and collaboration
|
| 112 |
|
| 113 |
+
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:
|
|
|
|
|
|
|
|
|
|
| 114 |
|
| 115 |
+
- **Email:** [contact@wirku.ng](mailto:contact@wirku.ng)
|
| 116 |
+
- **X (Twitter):** [https://x.com/FossDT](https://x.com/FossDT)
|
| 117 |
+
- **Web:** [https://wirku.ng](https://wirku.ng)
|
|
|
|
| 118 |
|
| 119 |
## Papers
|
| 120 |
|
| 121 |
+
- **[The Origami Model Family: Compiling Pretrained Foundation Models into Holographic Causal Graph Machines](paper/Foss_2026_Origami_HCGM.pdf)**. David Tom Foss, 2026. Included directly in this repository at `paper/Foss_2026_Origami_HCGM.pdf`.
|
| 122 |
+
- **[Holographic Causal Graph Machines: Source-Free Constant-State Execution of Pretrained Transformers](https://doi.org/10.5281/zenodo.19040842)**. David Tom Foss, 2026.
|
| 123 |
+
- **[Crystallized Attention: Constant-Memory Exact Store and Exponential-Mode Carriers](https://doi.org/10.5281/zenodo.19058017)**. David Tom Foss, 2026.
|
| 124 |
+
- **[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.
|
| 125 |
|
| 126 |
## Citation
|
| 127 |
|
| 128 |
```bibtex
|
| 129 |
+
@techreport{foss2026origami,
|
| 130 |
+
title = {The Origami Model Family: Compiling Pretrained Foundation Models into Holographic Causal Graph Machines},
|
| 131 |
+
author = {Foss, David Tom},
|
| 132 |
+
institution = {Wirkung Inc.},
|
| 133 |
+
year = {2026},
|
| 134 |
+
month = nov,
|
| 135 |
+
url = {https://wirku.ng}
|
| 136 |
}
|
| 137 |
```
|
| 138 |
|
| 139 |
## License
|
| 140 |
|
| 141 |
+
Released under the [Wirkung Model and Runtime License 1.0](LICENSE). Compiled in closed form without training steps from Upstream `MiniMax-H3` terms and notices (`NOTICE`); full attribution and closed-form transformation notices are recorded in [`NOTICE`](NOTICE).
|
assets/{weights_vs_origami.png → architecture.png}
RENAMED
|
File without changes
|
assets/architecture_organs.pdf
DELETED
|
Binary file (33.5 kB)
|
|
|
assets/banner.png
CHANGED
|
Git LFS Details
|
|
Git LFS Details
|
assets/{architecture_organs.png → mechanism.png}
RENAMED
|
File without changes
|
assets/{mrail_network.png → mrail.png}
RENAMED
|
File without changes
|
assets/mrail_network.pdf
DELETED
|
Binary file (29.6 kB)
|
|
|
assets/weights_vs_origami.pdf
DELETED
|
Binary file (46 kB)
|
|
|
paper/Foss_2026_Origami_HCGM.pdf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:72360f42140569826578d8c872b993a5be6ec1b43015d2536b82ec7e33ff9868
|
| 3 |
+
size 1716287
|