Spectral recurrence. Clifford rotors. Oscillator dynamics. Explore our open checkpoints and try the models in their native research demos.
Different sequence mechanisms, native byte vocabularies, and explicit checkpoint provenance.
An attention-free recurrent model with bounded state. The clean release completed one corpus pass, with no chat or instruction fine-tuning.
The spherical recurrent architecture with partial instruction tuning. Uses its native Instruction / Response format and a trained response terminator.
Benchmark-family-directed tuning; its scores are not a clean general-purpose evaluation.
A byte language model built around bivector coupling, Cayley rotors, and sphere-valued dynamics. The demo serves the Annealed mixed-domain checkpoint.
A small, from-scratch oscillator language model whose sequence mixing uses Clifford algebra and harmonic dynamics. Explore the tensors and configuration.
The public release contains weights and configuration; it has no bundled standalone inference runtime.
Conventional Transformer checkpoints help study behavior alongside the geometric architectures.
The 127.6M raw-byte chat model used as a conventional Transformer control. Sampling controls include temperature, top-p, top-k, and byte limits.
Dialogue represented as a room of peer agents, with explicit speaker, addressee, audience, and participant registry. A 127.6M Transformer dialogue experiment.
These are small experimental models. Outputs can be incorrect, repetitive, or malformed. Training and within-run metrics are documented on each model card; they should not be compared across different corpora or evaluation protocols.
Raw checkpoint archives preserve experiments and provenance. Supported inference releases are identified separately.