wflying/Qwen3.5-4B-RL-MATH

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Aug 12, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

wflying/Qwen3.5-4B-RL-MATH is a 4.5 billion parameter mathematical reasoning model, fine-tuned from Qwen/Qwen3.5-4B using reinforcement learning with verifiable rewards (RLVR) on a 15,000-example English mathematical reasoning dataset. This model, built on the Qwen3.5 multimodal conditional-generation architecture, is specifically optimized for competition mathematics and transferable reasoning in a non-thinking mode. It demonstrates significant improvements over its base model on benchmarks like AIME24, AIME25, and HMMT25, making it ideal for specialized mathematical problem-solving tasks.

Loading preview...

Overview

wflying/Qwen3.5-4B-RL-MATH is a specialized 4.5 billion parameter language model designed for advanced mathematical reasoning. It is post-trained from the Qwen/Qwen3.5-4B base model using a GRPO-style reinforcement learning with verifiable rewards (RLVR) objective. The training utilized the wflying/math-rlvr-15k dataset, focusing on improving performance in competition mathematics and general reasoning abilities.

Key Capabilities

  • Enhanced Mathematical Reasoning: Significantly improves performance on complex mathematical problems, as evidenced by substantial gains on benchmarks like AIME24 (+16.05 points), AIME25 (+20.21 points), and HMMT25 (+20.42 points) compared to its Qwen3.5-4B baseline.
  • Reinforcement Learning Optimization: Leverages a unique RLVR training method with rule-verifiable mathematical correctness signals to refine its problem-solving approach.
  • Transferable Reasoning: Designed to develop transferable reasoning skills, particularly in a non-thinking operational mode.

Good For

  • Competition Mathematics: Excels at solving problems typically found in mathematical competitions.
  • Specialized Mathematical Tasks: Ideal for applications requiring precise and verifiable mathematical solutions.
  • Research in RL for Reasoning: Provides a strong baseline for further research into reinforcement learning techniques for improving LLM reasoning capabilities.

Limitations

  • Task Specialization: Performance is optimized for mathematical reasoning; its effectiveness on unrelated tasks may vary.
  • Non-Thinking Mode: Evaluation results are based on a "non-thinking" mode, and performance may differ with other configurations or prompts.
  • Text-Only RLVR: While the base model is multimodal, the RLVR stage used only text-based mathematical data, meaning multimodal capabilities were not specifically optimized or evaluated in this post-training phase.
  • Potential for Incorrect Answers: Users should always verify generated mathematical solutions, especially in high-stakes environments.

Finetunes
1 models