LakshyAAAgrawal/QThink-Qwen3-1.7B-AIME2025

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Mar 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

LakshyAAAgrawal/QThink-Qwen3-1.7B-AIME2025 is a 1.7 billion parameter model based on Qwen3, developed by LakshyAAAgrawal. It is specifically fine-tuned for competition-level mathematics, utilizing a novel 'QThink' method for parallel latent reasoning via per-step distillation of multiple rollouts. This model is optimized for solving complex math problems, particularly those found in contests like AIME, despite being trained with a short context budget of 768 tokens.

Loading preview...

QThink-Qwen3-1.7B-AIME2025: Parallel Latent Reasoning for Math

This model, developed by LakshyAAAgrawal, is a 1.7 billion parameter Qwen3 variant fine-tuned for competition-level mathematics, specifically targeting AIME problems. It introduces the 'QThink' method, which employs parallel latent reasoning through per-step distillation of multiple rollouts.

Key Capabilities & Features

  • Specialized Math Reasoning: Trained on the DAPO-Math-17K dataset (14.1K problems) to excel in complex mathematical problem-solving.
  • QThink Method: Utilizes a unique distillation approach (uniform multi-rollout per-step distillation with gamma=2.0, K=6 latent steps) to enhance reasoning capabilities.
  • Cross-Benchmark Performance: Achieves strong results on various math benchmarks, including 83.2% on GSM8k, 43.6% on MATH-500, and 48.5% on Tooluse, outperforming SFT and base models in these areas.
  • Short-Context Training: Despite being trained with a limited 768-token budget (512 prompt, 256 answer), it demonstrates competitive performance, though a full-length version for AIME's longer reasoning chains is in progress.

When to Use This Model

  • Competition Math: Ideal for tasks requiring advanced mathematical reasoning, particularly those similar to AIME problems.
  • Research in Reasoning: Useful for exploring novel distillation and parallel reasoning techniques in LLMs.
  • Resource-Constrained Environments: Its 1.7B parameter size makes it suitable for applications where larger models might be impractical, while still offering specialized math capabilities.

Finetunes
1 models