trfrasdf/qwen3-32b-olympiad-cp

TEXT GENERATIONPricing:Input $0.408 / Cached $0.0816 / Output $1.972Concurrent Unit Cost:2Model Size:32BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 30, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

trfrasdf/qwen3-32b-olympiad-cp is a 32 billion parameter Qwen3-based model developed by trfrasdf, specifically fine-tuned for olympiad-level competitive programming. This model excels at step-by-step algorithmic reasoning and generating functional code for complex programming problems. It was trained using supervised fine-tuning and reinforcement learning with a verifiable reward based on passing hidden test cases, making it a specialized tool for code reasoning research.

Loading preview...

Qwen3-32B Olympiad-CP: Specialized Competitive Programming Reasoning Model

This model, developed by trfrasdf, is an open 32 billion parameter Qwen3-based model specifically fine-tuned for olympiad-level competitive programming. It is designed to reason through algorithmic problems step-by-step and produce complete, working code. It serves as a research artifact and a valuable starting point for those building or studying reasoning systems for code.

Key Capabilities

  • Algorithmic Reasoning: Excels at understanding and breaking down competitive programming problems.
  • Code Generation: Produces full programs based on the derived algorithmic reasoning.
  • Verifiable Training: Utilizes a unique reinforcement learning approach where rewards are based on whether generated code passes hidden test cases, ensuring objective performance signals.

Training Methodology

The model underwent a two-stage training process:

  1. Supervised Fine-Tuning (SFT): Applied to open code-reasoning datasets (RStar-Coder, OpenCodeReasoning) to teach step-by-step algorithmic reasoning and clean code output.
  2. Reinforcement Learning (GRPO): Employed a verifiable reward system where candidate solutions are run against hidden test cases, and the fraction of passed tests determines the reward. This stage used 8 NVIDIA B200 GPUs.

Intended Use Cases

  • Research into reasoning and reinforcement learning for code generation.
  • As a foundational model for further fine-tuning or development.
  • Generating and analyzing solutions to complex algorithmic problems.

Important Considerations

  • Specialized Niche: This is a specialized model, not a general-purpose assistant. It may not match the general coding capabilities of frontier models like GPT, Claude, or Gemini.
  • Verification Required: Code generated by the model should always be run and verified, as it can be confidently wrong.
  • No General Safety Tuning: Safety tuning is limited to that of the base Qwen3-32B model, as it is not intended for general assistant tasks.

Finetunes
1 models