Skip to content

Senter Ohm

The flagship model. See the full blog post: ../../blog/senter-ohm-flagship.md

Definition

Senter Ohm is the flagship model of the OmniSenter project — a ~32B-total, ~8B-active Mixture-of-Experts (top-1 routing) with the Ohm self-evolution engine bundled in.

The name decomposes: - Senter = Omni + agentic core (function calling, notebook management, plugin routing) - Ohm = the self-evolution engine (background CMA-ES, atomic weight swap, strict improvement acceptance)

The headline numbers
Total params ~32B
Active per token ~8B (top-1 routing)
Context window 256K (via YaRN)
Modalities text + vision + audio + video + music (in + out)
Self-evolution continuous, background, strict-acceptance
Inference VRAM ~22GB at 4-bit Q4_K_M, 32K context (fits 1× RTX 3090)
Disk (Q4_K_M GGUF) ~18GB
HF target sovthpaw/senter-ohm-32a8b (planned)

How it's built

The 5-stage pipeline (see ../../blog/the-5-stage-pipeline.md):

Stage What Input Output
1 Agentic SFT (QLoRA) Qwen3-8B senter-ohm-8b-sft
2 Evolutionary merge (CMA-ES) 3 specialized variants senter-ohm-8b-merged
3 Sparse upcycle to MoE merged 8B + 5 specialists senter-ohm-moe-32a8b
4 256K YaRN context extension MoE 32A8B senter-ohm-moe-32a8b-256k
5 Plugin + notebook + Ohm wiring 256K MoE deployable .ohm bundle

The 5-6 routed experts

  1. Agentic expert (function calling, tool use)
  2. Image/video expert (distilled from Qwen3-Omni-30B-A3B)
  3. Music expert (distilled from HeartMuLa / ACE-Step)
  4. Long-context expert (the YaRN-extended anchor)
  5. Synthesia expert (cross-modal memory, see synthesia.md)
  6. Generalist fallback (catches fall-throughs)

The 256K context

The 256K context is for the notebook — the structured state object that flows between Senter Ohm and Hermes Agent. The raw conversation stays short; the notebook entries get the long window.

See also