Skip to content

OmniSenter 12B — the small Senter

Status: ⏳ planned · HF target: sovthpaw/omnisenter-12b

Identity
Full name OmniSenter 12B
Type Small function-calling + omnimodal fusion (dense-ish)
Total params ~12B (aspirational)
Active per token ~12B (dense)
Context window 256K (target)
Modalities text + vision + audio + speech (in + out)
Self-evolution no (Senter, not Senter Ohm)

What it is

The small one. Built from Cosmos + Senter + Hermes-trained Qwen VL 8B Darwin children. The shipping target for routine function calling + omnimodal fusion on commodity hardware.

The "12B" is the aspirational active-parameter count — the actual merge may come out to ~10-14B active depending on how the Darwin children stack. It's a dense-ish model, not a MoE.

Build approach

  • Base: Cosmos3-Nano (multimodal reasoning)
    • Senter agentic core (function calling, notebook manager)
    • Hermes-trained Qwen VL 8B Darwin children (vision-language specialization)
  • All Darwin-merged via paper-exact 2-parent merge
  • Dense model, no MoE routing

Why this exists

The flagship (Senter Ohm 32A8B) is the moonshot. OmniSenter 12B is the shipping target — small enough to run on a single RTX 3090 in 4-bit (~6-8GB), fast enough for routine use, capable enough to handle function calling + notebook + multimodal I/O for typical agent workloads.

See also