PERSPECTIVE v2 · architecture research
OMNI
Where should the thinking happen?
An exploration of sparse experts, geometric routing and streamed execution under consumer hardware constraints. The architecture is ambitious; its unproven parts are explicit.
- State
- Partial implementation
- Working with
- Rust / C++ / sparse experts / recurrent state
01 The constraint
OMNI’s PERSPECTIVE v2 asks whether a large sparse model can be organized around limited memory and storage bandwidth. Rather than assuming every parameter stays in GPU memory, the design combines selective expert activation, layer streaming, low-precision weights and recurrent state.
02 The decisions
Make hardware part of the model.
The specification treats expert storage, transfer and execution as part of the architecture. Sparse activation reduces the work selected for a token; layer streaming addresses residency. Neither removes the need to measure the cost of moving and executing those weights.
Give routing a geometry.
Experts occupy a three-dimensional torus with wrap-around distance. The Rust manifold code implements positions, distances and neighbour relationships; routing updates fold new evidence into existing coordinates. The design uses locality to investigate neighbour-aware transfer.
Refuse a simulated success.
The token-processing path explicitly returns an error until execution is implemented. Component prototypes and tests can establish local behaviour, but they do not establish a trained language model. The roadmap separates those milestones and calls for smaller-scale validation first.
- 01Token state
- 02Manifold routing
- 03Streamed expert
- 04Updated state
03 What exists
The repository provides a Rust implementation workspace with routing, recurrent-state, memory, decoding, adaptation and safety components, alongside architecture and validation documents. Its most concrete result is an inspectable set of components and an explicit map of what still blocks end-to-end inference.
04 Still open
This is not a functioning production LLM. The token-processing path is incomplete; the end-to-end training system and GPU execution paths still require work. Parameter-scale and throughput figures in the specification are targets or projections, not demonstrated performance. Prototype safety geometry is not proof of general output safety.
Follow the evidence
Based on project code and documentation reviewed in October 2026. Evaluation results are reported by the projects; they were not independently rerun for this portfolio.
- Repository maturity and roadmap
Separates the specification, partial implementation and release requirements.
- Expert manifold
Implemented torus geometry, distances and expert neighbourhoods.
- Inference pipeline
The explicit fail-fast boundary in token processing.
- Validation plan
The checks required before architecture targets become evidence.