reaperdoesntknow/DualMind

Hugging Face
TEXT GENERATIONConcurrency Cost:1Model Size:2BQuant:BF16Ctx Length:32kPublished:Mar 29, 2026License:apache-2.0Architecture:Transformer Open Weights Warm

DualMind by Convergent Intelligence LLC is a 1.7 billion parameter Qwen3ForCausalLM model designed for dual-mental-modality reasoning. It employs a unique architecture where a single model uses role tokens (, , ) to simulate self-critique and refinement, enhancing logical inference. This model excels at structured problem-solving by internally generating, critiquing, and synthesizing solutions, making it suitable for tasks requiring robust reasoning and error detection.

Loading preview...

DualMind: Dual-Mental-Modality Reasoning

DualMind is a 1.7 billion parameter model from Convergent Intelligence LLC that introduces dual-mental-modality reasoning within a single Qwen3ForCausalLM architecture. It simulates an internal dialogue using role tokens to facilitate self-correction and refined problem-solving.

Key Capabilities

  • Structured Reasoning: Utilizes <explore> for unconstrained derivation, <examine> for adversarial self-critique, and <response> for clean synthesis of answers.
  • Self-Correction Mechanism: Provides a structural method for the model to detect and correct its own errors, mimicking the benefits of multi-model collision arrays within a single set of weights.
  • Enhanced Logical Inference: Trained on logical inference problems (initially KK04/LogicInference_OA and later Crownelius/Opus-4.6-Reasoning-3300x), restructuring CoT solutions into its unique cognitive loop format.
  • Qwen3-based Architecture: Built upon the Disctil-Qwen3-1.7B base model, featuring a 40,960 token context length and GQA attention.

Why Dual Modality?

Unlike standard Chain-of-Thought (CoT) prompting, DualMind's approach allows the model to freely explore solutions, identify potential flaws through an adversarial self-review, and then synthesize a more robust final answer. This process is grounded in Discrepancy Calculus, specifically Continuous Thought Dynamics, which models inference as a discrepancy-guided process where each phase plays a distinct role in minimizing error and refining output.