Model Overview

This model, developed by Convergent Intelligence LLC, is a 1.7 billion parameter Qwen3-based language model designed for structured reasoning. It was built using a unique two-stage training pipeline: first, knowledge distillation from a 30B MoE teacher (Qwen3-30B-A3B-Instruct-2507) to establish a robust STEM chain-of-thought reasoning backbone, and then supervised fine-tuning on legal instruction data to integrate domain-specific knowledge and instruction-following capabilities.

Key Capabilities

• Structured Reasoning: Employs a novel proof-weighted cross-entropy loss during distillation, prioritizing the learning of reasoning chains over memorizing answer formats, enabling the model to "learn how to think." This is further enhanced by KL divergence distillation at T=2.0 to capture the teacher's full probability landscape.
• Domain Transfer: Demonstrates effective transfer of structured reasoning patterns from STEM to legal domains, applying learned derivation skills to legal analysis rather than just memorizing templates.
• Dual Prompt Formats: Supports both a STEM derivation format (for rigorous step-by-step problem-solving) and an instruction-following format (for general questions and legal reasoning).
• Lightweight Deployment: At 1.7B parameters, it's suitable for edge/mobile deployment, with GGUF quantized versions available.

Good For

• Structured legal reasoning and analysis.
• STEM problem-solving requiring step-by-step derivations.
• Instruction-following across various technical domains.
• Educational tutoring and proof drafting.
• Integration into multi-model pipelines and retrieval-augmented workflows needing lightweight reasoning.