Jackrong/Qwen3-4B-Thinking-2507-GLM-4.7-Distilled

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Feb 23, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Jackrong/Qwen3-4B-Thinking-2507-GLM-4.7-Distilled is a 4 billion parameter Qwen3-based model developed by Jackrong, fine-tuned for structured Chain-of-Thought (CoT) execution and multi-turn conversational alignment. It adopts a "Plan-then-Execute" reasoning paradigm, systematically breaking down tasks and delivering modular, report-like responses. This model excels in mathematical reasoning and logical problem-solving, making it suitable for tasks requiring structured thought processes.

Loading preview...

Model Overview

Jackrong/Qwen3-4B-Thinking-2507-GLM-4.7-Distilled is a 4 billion parameter model built upon the Qwen3 architecture, developed by Jackrong. It is a highly specialized model focused on transforming reasoning patterns from a spontaneous, linear style to a structured "Plan-then-Execute" paradigm. This model was created through a two-stage training process:

  1. GRPO Reinforcement Learning: Starting from Qwen/Qwen3-4B-Thinking-2507, Group Relative Policy Optimization (GRPO) was applied using mathematical reasoning datasets like unsloth/OpenMathReasoning-mini and open-r1/DAPO-Math-17k-Processed. This stage significantly enhanced its mathematical problem-solving and step-by-step logical reasoning.
  2. SFT GLM-4.7 Distillation: The model was then fine-tuned using Supervised Fine-Tuning (SFT) with a mixed dataset heavily distilled from the GLM-4.7 model series (at a default temperature of 1.0). This distillation introduced greater lexical diversity, broader mode coverage, and more formatted Chain-of-Thought (CoT) traces, enabling the model to generalize across diverse conversational reasoning patterns and handle multi-turn conversations effectively.

Key Capabilities

  • Structured Chain-of-Thought (CoT): Transitions from a "think-as-you-go" approach to a "Plan-then-Execute" framework, breaking down tasks into logical outlines and executing modular responses.
  • Enhanced Reasoning: Improved correctness in mathematical problem-solving and logical reasoning due to GRPO training.
  • Multi-turn Conversational Alignment: Effectively handles complex multi-turn dialogues while maintaining internal CoT tracking.
  • Confident Execution: Exhibits high confidence in its outputs, rarely hesitating or expressing self-doubt, leading to objective and systematic responses.

Good for

  • Research and Education: Ideal for studying structured reasoning and mathematical problem-solving in LLMs.
  • Tasks requiring logical decomposition: Use cases where breaking down problems into clear, sequential steps is beneficial.
  • Applications needing verifiable reasoning traces: The model naturally generates <think>...</think> traces that can be parsed for inspecting its thought process.

Limitations

  • Intended primarily for research and educational purposes.
  • May still produce incorrect answers or hallucinations, requiring verification for critical tasks.
  • Not recommended for high-stakes applications without further safety evaluations.

Finetunes
1 models