reaperdoesntknow/DualMind

TEXT GENERATIONConcurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 29, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

DualMind by Convergent Intelligence LLC is a 1.7 billion parameter Qwen3ForCausalLM model with a 40,960 token context length, designed for dual-mental-modality reasoning. It employs a unique role-conditioned self-critique mechanism, using , , and tokens to simulate internal dialogue and self-correction. This architecture allows the model to derive, critique, and synthesize information, making it particularly effective for complex logical inference and problem-solving tasks.

Loading preview...

DualMind: Dual-Mental-Modality Reasoning

DualMind, developed by Convergent Intelligence LLC, is a 1.7 billion parameter model built on the Qwen3ForCausalLM architecture, featuring an extensive 40,960 token context length. Its core innovation lies in its dual-mental-modality reasoning system, which enables a single model to engage in a structured internal dialogue for enhanced problem-solving and self-correction.

Key Capabilities & Features

  • Role-Conditioned Generation: Utilizes specific role tokens (<explore>, <examine>, <response>) to guide the model through distinct cognitive phases:
    • <explore>: For unconstrained reasoning, derivation, and speculation.
    • <examine>: For adversarial self-response, critiquing its own explore output to detect errors and refine solutions.
    • <response>: For synthesizing a clean, final answer from the internal dialogue.
  • Self-Correction Mechanism: This dialectical structure allows the model to make mistakes and explore various approaches in the <explore> phase, then rigorously self-correct and verify in the <examine> phase, leading to more robust and accurate outputs.
  • Mathematical Foundations: Connects to Discrepancy Calculus (DISC) and Continuous Thought Dynamics, modeling inference as a discrepancy-guided process where the phases correspond to controlled trajectories through cognitive space.
  • Training: Fine-tuned on logical inference problems from the KK04/LogicInference_OA dataset, with an upcoming iteration on Crownelius/Opus-4.6-Reasoning-3300x for higher quality reasoning traces.

Good for

  • Complex Logical Inference: Excels at tasks requiring structured reasoning, proof generation, and problem-solving where self-correction is beneficial.
  • Enhanced Reliability: Ideal for applications where the accuracy and robustness of the generated output are critical, as the model actively critiques its own reasoning.
  • Research into Cognitive Architectures: Provides a practical implementation of dual-cognition within a single model, offering insights into advanced reasoning mechanisms.