Overview

DualMinded-Qwen3-1.7B is a 1.7 billion parameter model from Convergent Intelligence LLC: Research Division, built on the Qwen3 architecture. It features a novel dual-cognition approach, implementing a three-phase cognitive loop: <explore> for unconstrained reasoning, <examine> for self-critique and error detection, and <response> for synthesizing a clean answer. This dialectical structure is achieved through role-conditioned generation on shared weights, without additional parameters or routing.

Key Capabilities & Training

• Dual-Cognition Architecture: Employs an explore, examine, respond loop for enhanced reasoning and self-correction.
• Advanced Distillation: Trained using a four-stage pipeline including Multi-Teacher Distillation (from Qwen3-30B-A3B variants), Discrepancy Calculus (DISC) Refinement, and Topological Knowledge Distillation (TKD).
• Opus 4.6 Reasoning Traces: Fine-tuned on the Opus-4.6-Reasoning-3000x-filtered dataset, directly mapping thinking to <explore> and splitting solution into <examine> + <response>.
• Emergent Behaviors: Exhibits "ghost imprinting" from sequential distillation, leading to emergent capabilities not present in individual teachers.
• Mathematical Foundations: Grounded in Discrepancy Calculus, a measure-theoretic framework that preserves structural boundaries during knowledge distillation.

Good For

• Tasks requiring extended reasoning and creative derivation.
• Applications benefiting from self-critique and refined outputs.
• Scenarios where a small, efficient model (1.7B parameters) needs advanced reasoning capabilities.
• Exploring models trained with novel distillation techniques like DISC and TKD.