About OmniSenter¶
TOWARDS SELF-IMPROVEMENT
What is OmniSenter?¶
OmniSenter is a project to build a multi-stage agentic MoE with the Ohm self-evolution engine — and to make it a useful auxiliary to the Hermes Agent.
The model that the project produces is called Senter Ohm — a ~32B total / 8B active sparse-upcycled Mixture-of-Experts with the self-evolution engine bundled in. The 256K context window is for the notebook — the structured state object that flows between Senter Ohm and Hermes.
The three load-bearing ideas¶
- Darwin Family weight-space recombination (after Akiba et al.'s Evolutionary Optimization of Model Merging Recipes, 2024) — a training-free way to merge LLMs that outperforms either parent.
- Sparse upcycling (per Komatsuzaki et al. 2022) — turn a dense model into a MoE by copying the FFN as N parallel experts + a router. Cheap, fast, and the active compute stays the same.
- The Ohm runtime — the model file is the engine. Background CMA-ES loop, atomic weight swap, strict-acceptance policy. The model gets a tiny bit better every 5 minutes while it serves.
The naming¶
- Omni = multimodal native (text + vision + audio + video + music)
- Senter = Omni with the agentic core wired in (function calling, the notebook, plugin routing)
- Ohm = the self-evolution engine (the
.ohmfile format) - Senter Ohm = the flagship — all three composited (~32A8B MoE)
- OmniSenter = the project itself, and the small Senter
(
OmniSenter 12B)
The team¶
OmniSenter is a project by Chris (sovthpaw), built at Nous Research. The blog is written in collaboration with Omni VA (the Senter-side voice).
The repos¶
The OmniSenter system spans 9 GitHub repos:
| Repo | Role |
|---|---|
evolutionary-training |
Main repo, training scripts, Ohm runtime |
evolutionary-model-merging |
Darwin Family (CMA-ES + paper-exact merge) |
multimodal-expansion |
REAP + EvoMoE + sparse_upcycle |
omnistep-fusion |
Cosmos × ACE-Step multimodal merge |
evolutionary-radio |
The OmniStep-brained music radio |
hermes-agent |
The smart agent Senter is auxiliary to |
senter |
Senter Hermes profile |
nous-girl |
Omni VA Hermes profile |
chizul |
Chizul Hermes profile |
The HuggingFace models¶
| Model | Size | Status |
|---|---|---|
sovthpaw/omnistep-12a3b |
12A3B | ✅ published (transitional) |
sovthpaw/Omni-Senter-3B |
3B | ✅ published (transitional) |
sovthpaw/OmniSenter-Base-16B |
16B | ✅ published (transitional) |
sovthpaw/omnisenter-12b |
~12B | ⏳ planned |
sovthpaw/omnisenterstep |
TBD | ⏳ planned |
sovthpaw/senter-ohm-32a8b |
~32A8B | ⏳ planned |
The skill stack¶
Built on:
- TRL for SFT
- PEFT for QLoRA
- bitsandbytes for 4-bit
- llama.cpp for inference
- Xurl for X integration
- MkDocs Material for this site
License¶
Apache 2.0 (following parent model licenses).
TOWARDS SELF-IMPROVEMENT