Giving Reasoning Capabilities to Small Models: Proven Facts vs Real Limitations

Published on September 10, 2026 · Transparent assessment of Alice lightweight reasoning architecture: 511 tokens/s on CPU, zero subscription, and real limitations.

Dual Architecture: Local CPU Reflexes & Supervision Nano-Reasoner CPU 511 tokens/s · 4.68 MB CPU L3 Cache · 0 € subscription Step-by-step logical deduction & audit 6-Agent Orchestration Rule Arbitration & Firewall Sub-250ms decisions Selective delegation to large models

The artificial intelligence industry is realizing the dead-end of cloud-only architectures: giant models are slow and costly for ground-level tasks that demand speed and frugality. With the Alice project, we developed and tested a lightweight reasoning architecture executed directly on consumer hardware. Here is the exact assessment: what is formally measured and proven, and the system's actual limitations.

1. What is Measured and Proven (The Facts)

2. What is NOT Proven and Actual Limitations

An honest engineering approach requires clearly stating what the system cannot do to avoid any magical thinking.

3. Why This Architecture Changes the Game

This result proves that an autonomous agent does not require a supercomputer to be effective. By combining a live dynamic map, procedural memory, and a dedicated micro-reasoning model, a modest home computer becomes a dependable, auditable, and instantaneous copilot.