Models
Research into perception, context, and model computation.
Theia
Adaptive video tokenization for agentic perception.
Theia changes the visual token rate according to how much a video is changing. Static frames receive few or no tokens; action-driven transitions receive more. The current prototype implements and tests the core tokenizer components.
- Status
- Research prototype
- Base model
- NVIDIA Nemotron 3 Nano Omni
Hyperion
Context self-rewriting for memory management and speculative thinking.
Hyperion asks whether a language model can rewrite its active context during inference. It is currently a research concept, with no selected base model, executable architecture, or measured results.
- Status
- Research concept
- Base model
- Not selected
Iapyx
Self-modifying computation graphs for frozen MoE language models.
Iapyx keeps a Mixture-of-Experts model frozen while a small trainable head proposes changes to its forward pass. The prototype can override expert routing, loop states through earlier layers, dispatch validated kernels, and snapshot or roll back speculative changes.
- Status
- Research prototype
- Base model
- Qwen3-30B-A3B