# COSMOS (54D)

This model is COSMOS — a 54D transformer architecture packaged as a GGUF for local runtimes. The GGUF hosted here is the patched COSMOS architecture verified on Ollama.

## Quick Links
- **[Performance Profile](performance.md)** — Latency, memory, and throughput metrics
- **Weights:**  1_HER_SOUL/weights/cosmos-namebind-weights.gguf (3.1 GB Q4)
- **Backup:**  1_HER_SOUL/weights/cosmos-namebind-weights-BACKUP-before-replace.gguf

## Architecture
COSMOS combines:
- **12D CST Phase Attention** — Geometric phase-space attention
- **24D Hebbian Plasticity** — Online synaptic weight adaptation during inference
- **18D Chaos Oscillators** — Deterministic Lorenz dynamics to prevent collapse
- **256-slot Persistent Memory** — True session continuity

## Compatibility notes
- Runtime tensor layout may be qwen2-compatible; file metadata expresses COSMOS architecture. Ollama accepts this patched file.
- If other runtimes fail, use the BACKUP file in the repo or re-export from checkpoints.

## Quick start (Ollama)
1. Download weights: \ 1_HER_SOUL/weights/cosmos-namebind-weights.gguf\
2. Create Modelfile: \FROM .\\cosmos-namebind-weights.gguf\
3. \ollama create cosmos -f Modelfile\ then \ollama run cosmos "Hello"\

## Performance
- **Multi-turn latency:** ~5.9 s average per response
- **Memory footprint:** ~24–25 MB (CLI process peak)
- **Concurrency:** Tested with 3 parallel requests, all successful

See [performance.md](performance.md) for detailed benchmarks.

## Safety & Limitations
- Online Hebbian plasticity is a novel feature; careful validation recommended for production use
- Chaos oscillators may introduce stochasticity in edge cases
- Backup original GGUF for runtime compatibility fallback

---

**Status:** Tested and verified on local Ollama (Windows). Ready for broader platform deployment.