gasolsun/Math_Qwen3-1.7B_Qwen3-4B-Non-Thinking-RL-Math-Step500

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 15, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Math_Qwen3-1.7B_Qwen3-4B-Non-Thinking-RL-Math-Step500 is a 2 billion parameter causal language model, distilled from Qwen3-4B, specifically optimized for mathematical problem-solving. It leverages an innovative "Ask to Distill" method, training a question generator with reinforcement learning to create a synthetic question bank for on-policy distillation. This model excels in mathematical reasoning, achieving an average score of 46.16 across six math benchmarks, including AIME and MATH500, and operates in a non-thinking mode for direct problem solving.

Loading preview...

Overview

Math_Qwen3-1.7B_Qwen3-4B-Non-Thinking-RL-Math-Step500 is a 2 billion parameter language model based on the Qwen3-1.7B architecture, specifically distilled for enhanced mathematical problem-solving. It was created using an innovative technique called Ask to Distill, which involves training a question generator with reinforcement learning to produce a synthetic question bank. This method allows for on-policy distillation (OPD) without relying on real training questions or reference answers, focusing on token-level disagreement between a teacher (Qwen3-4B) and the student model.

Key Capabilities

  • Advanced Mathematical Reasoning: Specialized in solving complex math problems step-by-step.
  • Efficient Distillation: Utilizes the "Ask to Distill" method for effective knowledge transfer from a larger teacher model (Qwen3-4B) to a smaller student model (Qwen3-1.7B).
  • Non-Thinking Mode: Designed and evaluated to operate efficiently in a direct, non-thinking problem-solving mode, indicated by enable_thinking=False in its chat template.

Performance

The model demonstrates strong performance on various mathematical benchmarks:

  • AIME24: 34.8
  • AIME25: 26.0
  • AIME26: 26.5
  • AMC23: 71.9
  • BRUMO25: 35.4
  • MATH500: 82.4
    It achieves an average score of 46.16 across these six benchmarks (avg6), evaluated at temperature 0.7 and top-p 0.95.

Good for

  • Mathematical Problem Solving: Ideal for applications requiring accurate and step-by-step solutions to math problems.
  • Resource-Constrained Environments: Its 2 billion parameter size makes it suitable for deployment where computational resources are a consideration, while still offering strong mathematical capabilities.
  • Research in Distillation Techniques: Provides a practical example of the "Ask to Distill" method for model training.

Finetunes
1 models