sovthpaw¶
A curated stack of local models, self-evolving agents, and ever-better radio.
Built by sovthpaw (Chris). Forged on dual RTX 3090s in the long quiet hours.
OMNISENTER
Multimodal · Agentic · Self-Evolving
What lives here¶
This is the front door. Everything I build lands somewhere you can reach from this page.
| Section | What's in it |
|---|---|
| Blog | Long-form posts on the OmniSenter pipeline, the assistant, the radio, sparse upcycling, the Synthesia layer, the Ohm engine. |
| Projects | The repos. Omni VA, OmniSenter, the assistant, the radio, model merging, evolutionary training. |
| Models | The hand-curated catalog of local models and API combos. |
| Wiki | The reference. Every concept and entity, cross-linked. |
| About | Who I am, what I'm doing, and why. |
The Omni Family¶
A naming convention that ties together everything I build:
- Omni — multimodal native (text + vision + audio + video + music)
- Senter — Omni with the agentic core wired in (function calling, tool use, the notebook)
- Ohm — the self-evolution engine (CMA-ES background loop, atomic weight swap, strict improvement acceptance)
- Senter Ohm — the flagship ~32B-total / 8B-active MoE with all three composited
The flagship¶
Senter Ohm is a ~32B total / 8B active sparse-upcycled MoE with the Ohm self-evolution engine bundled in. The 256K context window is for the notebook — the structured state object that flows between Senter Ohm and Hermes Agent.
| Total params | ~32B |
| Active params | ~8B (top-1 routing) |
| Experts/layer | 5-6 |
| Context | 256K (after YaRN) |
| Modalities | text, voice, music, vision |
| Evolution | Ohm chain runs nightly |
The assistant¶
Omni VA is the desktop assistant that serves as the user-facing surface for the whole stack. She runs locally, is multimodal when OmniStep is the model, and falls back gracefully to Edge TTS when you swap in a text-only LLM. The Evolution Radio is bundled — it runs alongside whatever model is loaded, curates your taste, and feeds the Ohm chain.
The agent stack¶
Three tiers, all one-way write:
Omni VA (curation) → Senter (prioritization) → Hermes main (execution)
↑ ↓
└────────── writes notes to ~/wiki/assistant-curated/ ────────────┘
- Omni VA is the bounded curator — web, fetch, notes, social. No code, no terminal, no delegation.
- Senter is the on-demand triage — reads the wiki, returns a ranked list, defers execution.
- Hermes main is the doer — full toolset, reads the wiki, acts on the curated ideas.
From the blog¶
Flagship post — Senter Ohm: The Self-Evolving Flagship A ~32B total / 8B active MoE with the Ohm self-evolution engine. Three new ideas, one architecture diagram.
Cross-modal memory — Synthesia: The Cross-Modal Memory Layer Joint (text, audio, image) embeddings, 10 concrete benefits, the data it needs, the MoE expert.
The integration pattern — Senter as the Hermes Auxiliary How Senter talks to Hermes Agent. The notebook-as-API pattern, escalation rules, the cost model.
Live now¶
- Public site (you are here) — southpawin.github.io
- Pet + radio: nous-girl-agent
- Training pipeline: evolutionary-training
- Model merging: evolutionary-model-merging
- Multimodal expansion: multimodal-expansion
- HuggingFace models (v1 transitional): sovthpaw
Currently¶
- Stage 1 SFT running on Qwen3-8B base, 8B QLoRA, 4268 steps. ~step 1000 complete, loss 0.3959.
- Synthesia layer concept documented. Cross-modal memory indexer.
- Ohm engine designed. Atomic weight swap, CMA-ES background loop.
- 32A8B MoE math validated. 5-6 experts/layer, top-1 routing.