ZemoZemo
Zemo

An ACC Neural Cognition Engine

A conversation engine built on one pretrained base and shipped in three configurations. The difference between them is not the prompt but the shape of the network itself.

Members only

Three engines

Switch between them inside the chat at any time.

CorePhiModel

The PhiModel configuration. Pretrained weights are reshaped to golden-ratio relationships, so the pattern lives in the network's own structure rather than in how it is asked to behave.

  • MoE Transformer + PhiModel
  • 55 layers · 377 routed experts / layer
  • ~0.92T total · ~29.5B active
TurboPruned

The pruned configuration. Standard network pruning brings it to Core's structural counts, with no PhiModel reshaping applied at any layer.

  • MoE Transformer
  • 55 layers · 377 routed experts / layer
  • ~0.92T total · ~29.5B active
InfiniteAgentic

The furthest configuration. A fully agentic Zemo running well beyond the other two, with internal systems and technical specifications that are not available on any other model here.

  • Fully agentic runtime
  • Distinct internal configuration
  • Specifications undisclosed

Specifications

Every configuration starts from the same pretrained base. Structural counts are matched across Turbo and Core, so PhiModel reshaping is the only architectural difference between them.

SpecificationBase (Kimi K2.5)TurboCore
Architecture
Model typeMoE TransformerMoE TransformerMoE Transformer + PhiModel
Total layers615555
Dense layers111
MoE layers605454
Routing
Routed experts / MoE layer384377377
Shared experts / MoE layer111
Experts per token888
Scale
Total parameters~1.04T~0.92T~0.92T
Activated parameters~32B~29.5B~29.5B
Hidden dimension7,1687,1687,168
Expert FFN dimension2,0482,0482,048
Attention & context
Attention heads646464
Attention mechanismMulti-head Latent AttentionMulti-head Latent AttentionMulti-head Latent Attention
Activation functionSwiGLUSwiGLUSwiGLU
Context window256K tokens196,418 tokens196,418 tokens
Input
Vision encoderMoonViTMoonViTMoonViT
ModalitiesText, imagesText, imagesText, images
Vocabulary~160K~160K~160K
Serving
Serving stackSharedSharedShared
Response delayMatchedMatchedMatched
InfiniteAgentic

Infinite sits above both configurations. It runs a fully agentic architecture with its own internal systems and technical specifications that are not available on any other model here. Its architecture, parameter budget and internal state handling stay private, because the configuration itself is the work. What is public is the behaviour.

Configuration
Fully agentic
Runtime
Independent
Internal systems
Not published
Parameter budget
Undisclosed
Latency profile
Matched to Core and Turbo
Availability
Session based

Values marked with ~ are estimates derived from structural counts.

Images

Zemo answers with reaction images from its own library when words are not enough.

Reactions

It can drop an emoji on anything you say, chosen by the tone of the moment.

Presence

Zemo types as it thinks, so a reply lands at the pace a person would send it.

Have a passkey?

Zemo is open to ACC members. Sign in with your passkey and start a conversation.