Samsung TRM: 5–7M Params, 87.4% Sudoku-Extreme, Beats DeepSeek-R1

Samsung TRM: 5–7M Params, 87.4% Sudoku-Extreme, Beats DeepSeek-R1

Samsung's Tiny Recursive Model achieves 87.4% on Sudoku-Extreme with just 5–7 million parameters — compared to HRM's 55% and DeepSeek-R1's 0% on the same benchmark. On ARC-AGI-1 it scores 45%. TRM strips the Hierarchical Reasoning Model down to a single weight-shared two-layer network, proving the recursive loop itself — not biological hierarchy — is the core driver of deterministic reasoning gains.

Why It Matters

This validates recursive architectures as a cost-efficient path to deterministic reasoning, offering potential 100× speedup on such tasks versus equivalent LLM approaches. Embodied AI and latency-sensitive scientific computing become significantly more viable at these parameter counts.