ECHELON: How Modular Reasoning Cores Enable Small Models to Gain Advanced Reasoning Capabilities

ECHELON proposes a disruptive AI reasoning architecture, whose core is separating the executor (small frozen model) from the reasoning core (pluggable cards), realizing the vision of "carry the least, win the most". It addresses issues such as high cost, poor flexibility, and black-box trust of large models. Through an honesty mechanism (skills must be obtained through actual verification), it ensures credibility, and its effectiveness has been verified by empirical evidence (a 668-byte core enables gemma-4B to complete complex reasoning). The project adopts a dual-license model, with open source (AGPL) and commercial versions running in parallel.

The current "bigger is better" paradigm of large models has fundamental problems:

1. Cost issue: Each inference requires paying for the complete context, regardless of actual needs;
2. Flexibility issue: After the model is frozen, learning new skills requires expensive fine-tuning or retraining;
3. Trust issue: The internal working principle is black-boxed, making it difficult to verify the honesty of the reasoning process. ECHELON is an alternative solution targeting these pain points.

Core Idea: Separate "execution" and "reasoning"—the executor is a small frozen model (console), and reasoning capabilities come from external composable reasoning cores (cores). Technical Architecture:

• Atom: Learned parameter values that gain rankings through actual use;
• Card: A reasoning layer composed of atoms, which can be nested;
• Chain: The connection relationship between cards, defining the reasoning path. Honesty Mechanism: Skills must be verified through "earn-by-trace"—credit from successful tasks is backpropagated to strengthen components, while no credit is given for failures, ensuring component capabilities are traceable.

Under the same gemma-4B model, prompts, and decoding parameters:

• Without ECHELON: Failed on 4-rule chain problems (missing steps);
• With ECHELON: Successfully gave accurate answers via a 668-byte core (containing 4 atoms + 4 chain cards). Compared to fine-tuning (which requires megabytes of data, days of training, separate training, and is prone to forgetting), ECHELON is more efficient and flexible.

ECHELON opens up new directions for AI applications:

• Edge Computing: Small devices run frozen executors, and specific reasoning cores are downloaded from the cloud to achieve on-demand intelligence;
• Enterprise Scenarios: Sensitive data remains locally processed, and general reasoning capabilities are provided by standardized cores;
• Auditability: Reasoning steps can be traced and verified, providing a technical foundation for compliance and regulation.

ECHELON adopts a dual-license model:

• Open Source Version: GNU AGPL-3.0, allowing free use, modification, and sharing, but web services built on it must be open source;
• Commercial Version: Allows use in closed-source/proprietary products without complying with AGPL obligations. The project is developed by independent researcher Albert Tenggono, reflecting the value of individual innovation.

Current Limitations: Only small-scale mechanism verification has been completed, and full autonomy (automatically selecting and traversing the correct chain) has not yet been achieved. Future Directions: Reuse the scheduler on the wired card graph to realize autonomous system decision-making. The author's candor about limitations embodies ECHELON's "honesty" concept.

ECHELON represents an alternative to the mainstream path of large models—it does not pursue larger models or longer contexts, but enables small models to complete complex reasoning through architectural design. "Carry the least, win the most" is not only a technical slogan but also a concept of sustainable AI development. In today's era of tight computing resources and growing concern about environmental impact, it may be a more responsible technical direction.