Skip to content

README.md

Latest commit

 

History

History
90 lines (63 loc) · 2.17 KB

README.md

File metadata and controls

90 lines (63 loc) · 2.17 KB

Nexus Fast Mini

Nexus Fast Mini is a multi-model reasoning framework built around a simple idea: combine several specialized models, collect their perspectives, synthesize them, and produce a single high-quality answer.

This project provides the full TypeScript implementation of the Nexus pipeline, including:

  • Seven independent “thinker” models
  • A condenser stage that merges their insights
  • A chief reasoner that generates the final response
  • Streaming and non-streaming output modes
  • A clean developer API for integration

Nexus does not train or host models. Instead, it orchestrates external LLMs through the OpenRouter API.

How It Works

1. Thinkers

Nexus runs 7 models in parallel. Each one plays a specific role such as:

  • Strategic analysis
  • Critical review
  • Creative alternatives
  • Technical evaluation
  • UX considerations
  • Data-driven reasoning
  • Optimization and simplification

Each thinker produces a structured analysis.

2. Condenser

The condenser receives all thinker outputs and creates:

  1. Consensus points
  2. Disagreements
  3. A merged synthesis
  4. A short list of recommended actions

This stage reduces multiple viewpoints into a single brief.

3. Chief Reasoner

The chief takes the condensed summary and writes the final answer. It supports several output styles:

  • concise
  • detailed
  • structured
  • writing
  • coding

The chief produces the final, unified response.

Model Selection

The default models are defined at the top of the file. You are free to replace them with any models supported by OpenRouter.

Each thinker can be customized with:

{
  model: string;
  role: string;
  prompt: string;
}

Design Goals

  • Straightforward code that is easy to read and modify
  • No hidden logic or opaque abstractions
  • Fully deterministic structure with clear model roles
  • Works with any OpenRouter-compatible endpoint
  • Suitable for experimentation, extensions, or academic use

Notes

  • This project does not include any model weights.
  • All calls depend on the user’s OpenRouter key and quotas.
  • For custom deployments, modify or replace the OpenRouter client instance.