Wirkung says its Origami family turns five existing AI models into compact files that can run from disk, including a 78-billion-parameter model on a standard CPU server. If the company’s figures hold up, this could make capable models easier to run on ordinary hardware, rather than keeping them tied to large GPU systems.
Watch Desk analysis
What happened
In a Hugging Face community article, David Tom Foss says Wirkung compiled five pretrained models without retraining them. The resulting files range from 2.74 GB for a 9-billion-parameter model to 26.42 GB for the 78-billion-parameter Origami-78B. The post says the runtime needs no CUDA, vLLM, MLX or Transformers, and uses between 253 and 480 MiB of resident memory.
Wirkung also claims that Origami-78B runs on a 16-core CPU server at 9.80 tokens per second, and on a 16 GB Mac mini at 7.30. Its post describes Origami as a preview and links to a technical report.
Why it matters
Running a large model without a GPU would change who can host it and what hardware they need. The claimed constant-size session state could also ease one of the familiar constraints on long conversations: memory use growing with context. Wirkung says its 78B model held 0.62 GB of state across contexts from one million to 5.24 million tokens.
Those are the company’s results, not independently established performance figures. Still, this is a substantial technical claim with practical consequences, not merely a smaller download badge.
Our read
The interesting promise is not simply compression. Wirkung says it has changed how these models execute, while preserving outputs exactly across several hardware types. If that combination survives outside the company’s own tests, local AI could become a more realistic option on modest machines. The next useful step is independent testing against the original models, especially for quality, speed and compatibility.
What to watch
- Whether the technical report provides enough detail for others to reproduce the results.
- How Origami performs on independent tests against each original model.
- Whether the preview files and runtime become available for developers to try.
Discussion spark: If independent tests confirm these results, would you rather run a large model locally on modest hardware, or use a conventional GPU-backed service?
Sources and evidence
- Holographic Causal Graph Machines: The Origami Model Family (12 November 2026, 02:13 UTC)
- tomaarsen (Tom Aarsen) (12 November 2026, 02:13 UTC)
Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.