Physarum micro-model solves ARC with 250k parameters

Physarum Explorer proves that reasoning depth in AI can be achieved through bio-mimetic efficiency rather than just massive parameter scaling.

• –The use of "structural fingerprints" allows the model to recognize geometric invariants without the overhead of traditional computer vision.
• –Running at 8-11 FPS on a CPU enables high-frequency "thinking" loops that are currently too expensive for large-scale reasoning models.
• –The "ENGRAM" fuzzy memory system effectively bridges the gap between short-term exploration and long-term logical pattern matching.
• –While successful at initial levels, the model's struggle with higher-complexity tasks highlights a persistent "reasoning ceiling" for purely search-based architectures.
• –This approach suggests a path forward for "edge-AI" reasoning agents that don't require massive GPU clusters to perform abstract induction.