Fathom-R1-14B: Cost-Efficient Math Reasoning

FractalAIResearch's Fathom-R1-14B is a 14-billion-parameter language model specifically engineered for advanced mathematical reasoning. Derived from Deepseek-R1-Distilled-Qwen-14B, this model achieves state-of-the-art performance on olympiad-level math exams (AIME-25, HMMT-25) within a 16K token context window, rivaling closed-source o4-mini models. Notably, its post-training was achieved at an exceptionally low cost of $499, utilizing a supervised fine-tuning (SFT) approach on carefully curated datasets.

Key Capabilities

• Superior Mathematical Reasoning: Achieves 52.71% Pass@1 on AIME2025 and 35.26% Pass@1 on HMMT25, outperforming its base model and other open-source 14B models.
• Cost-Effective Training: Developed with a total post-training cost of just $499, demonstrating high efficiency in achieving advanced reasoning capabilities.
• Optimized for 16K Context: Designed to maximize performance within a 16K context window, balancing accuracy with practical deployment constraints.
• Generalization Beyond Math: Shows out-of-domain improvement on the GPQA-Diamond benchmark, suggesting broader reasoning capabilities despite math-centric training.
• Token Efficiency: Fathom-R1-14B-RS variant uses fewer response tokens than LightR1-14B on AIME25 and HMMT25 while maintaining higher Pass@1 scores.

Good for

• Developers and researchers focused on mathematical problem-solving and reasoning tasks.
• Applications requiring high accuracy on complex math challenges like those found in olympiads.
• Use cases where cost-efficiency and performance within a constrained context window (16K tokens) are critical.
• Exploring generalization of reasoning abilities from specialized mathematical training to other domains.