What it is

Apex is a semi-autonomous MoE-style reasoning prototype, designed to handle multiple-tasks along with a persistent internal state.

It explores how modular agentic systems can coordinate reasoning, decision making and adaptation without preset, rigid pipelines.
It's currently a 2x120B+8x8B architecture, each model orchestrating and managing varied tasks to make Apex the complete package, managing internal-state awareness, a functional DMN, and a custom module to simulate neuroplasticity.

Why I'm building it

I am building Apex to understand how complex systems like these evolve over time, how players like OpenAI and Anthropic are currently solving these issues, and what the extent of our understanding of agentic systems is today.

It is an attempt to solve some of these problems, to build a system that could reason, evaluate, enact, and learn on its own with minimal and abstract external help only.

How it works(High-Level)

At a conceptual level, Apex is built around modular components that interact through a shared controller.

• Modular reasoning units for task-specific processing(Coding, Philosophy, Brainstorming, etc)

• Dynamic task routing based on GPToss-like calling and mixture-based control

• Internal state tracking to preserve context across steps

• Iterative feedback loops and LoRA training/re-training, similar to MIT's SEAL architecture.

• DMN and introspection loops for memory processing, optimization, and pruning.

• Memory controller and automatic retrieval from different memory types(semantic, episodic, procedural, etc.