Website · Origami Family Paper · Runtime (PyPI) · Prior Constant-State Releases · X (@FossDT) · Contact

Wirkung builds zero-retraining model compilers, lossless and spectral weight-folding formats, constant-memory (O(1)) attention carriers, and zero-VRAM disk-mapped inference runtimes across CPU, NPU, and GPU. Every system follows one decomposition, X = X_known + R: keep what is known as an exact program and compute only the unseen remainder.

pip install -U wirkung && wkng pull wirkung/Origami-9B -o ./Origami-9B && wkng verify ./Origami-9B

Origami Model Family (Preview)

Every Origami Preview model is compiled in closed form from its pretrained upstream checkpoint with 0 retraining steps, 0 fine-tuning steps, and 0 gradient updates. Weights execute directly from SSD/NVMe storage via read-only OS mmap (0 GB GPU VRAM and 0 GB dedicated weight-RAM required) with a strictly bounded O(1) session state across arbitrary context lengths.

Repository	Architecture & Modality	Parameters	Source Checkpoint vs. Folded Bundle	Live Horizon State
`wirkung/Origami-78B`	Text & code MoE (`Aleph-Alpha/Kolibri-1`, 19,200 experts)	78.10B (3.46B active)	156.21 GB BF16 → 26.42 GB `.o1f1` (5.91×, SHA-256 lossless round-3)	0.183 GB constant `O(1)` carrier
`wirkung/Origami-31B`	Multimodal dense (`google/gemma-4-31B-it`)	32.68B	62.55 GB BF16 → 11.56 GB `.o3f` (5.41×)	68.40 MB constant `O(1)` carrier
`wirkung/Origami-27B`	Multimodal dense + rescued MTP (`Qwen/Qwen3.5-27B`)	27.80B	55.56 GB BF16 → 7.44 GB `.o3f` (7.46×)	34.24 MB constant `O(1)` carrier
`wirkung/Origami-9B`	Multimodal GDN + attention (`Qwen/Qwen3.5-9B`)	9.41B	18.85 GB BF16 → 2.74 GB `.o3f` (6.88×)	14.82 MB constant `O(1)` carrier
`wirkung/Origami-H3`	Audio-video diffusion DiT (`MiniMax-H3`, 50 DiT + 3D/Audio VAE)	21.60B	44.89 GB BF16 → 10.38 GB `.o3f` (4.33×)	Constant `O(1)` uncut horizon state

First-generation constant-state model releases compatible with standard vLLM, MLX, and Transformers stacks remain available under our sister organization huggingface.co/tfwnotops.

Core systems

Holographic Causal Graph Machine (HCGM) and .o3f / .o1f1 folding. Turns passive checkpoints into self-contained binary files on disk without retraining. Under the v3-exakt dyadic integer reduction contract, AVX2, NEON, general NPUs, sGPU, Apple Metal, and standalone driver-API PTX evaluate identical canonical reductions from a single artifact.
Wirkung Engine. Internal exact structural-law and closed-form operator synthesis engine used to derive constant-memory attention carriers (ckv3, zero-bit key-mean centering, exact doubly-stochastic and Sinkhorn attention surrogates) and compiled algebraic recurrences directly from empirical operators.
Shared M-Rail (.mrail v3) and GPU-less server fleets. Content-addressed associative trajectory manifold shared across processes or enterprise networks. Recurring workflows can be warmed up on one GPU (or a short cloud lease) and served permanently on disk-only CPU/NPU fleets; any unseen residual (R ≠ 0) solved once by a single node is committed immediately for the entire fleet.
Heavy Mixture of Experts (HMoE). Deterministic routing across multiple disk-resident Origami models, static-graph NPU specialist pools, and sub-2 KB exact algebraic organs on a synchronous station bus (cstar) at zero idle weight-RAM cost.

Inquiries, custom compilation, and research collaboration

We are available for custom zero-retraining model compilation (HCGM and ConstantKV), exact formula and algorithm synthesis via Wirkung, Shared M-Rail fleet licensing, enterprise partnerships, and research inquiries: