News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Fathom-R1-14B

Fathom-R1-14B

Low-cost open reasoning model for mathematics

Fractal AI Research released Fathom-R1-14B, a 14-billion-parameter language model focused on mathematical reasoning. Derived from DeepSeek-R1-Distilled-Qwen-14B, it was post-trained for 499 US dollars and operates within a 16,000-token context window.

Share on WhatsApp

  • Post-training cost of only 499 US dollars
  • Surpasses o1-mini and o3-mini-low on math benchmarks
  • Operates within a 16,000 token context window
  • Achieves 76.7% consistency on AIME 2025 exam
  • Open-sourced weights, recipes and datasets

What it is

Fathom-R1-14B is a specialised language model designed for complex mathematical problem solving. Developed by Fractal AI Research, it contains 14 billion parameters and builds upon the DeepSeek-R1-Distilled-Qwen-14B base. The model aims to deliver high-level reasoning capabilities similar to proprietary systems like o4-mini, but at a fraction of the computational cost. It operates within a constrained context window of 16,000 tokens, making it suitable for environments with limited memory resources. The project highlights how targeted post-training can enhance performance without requiring massive infrastructure budgets.

How it came about

The development process involved creating several intermediate versions before arriving at the final release. Researchers first produced Fathom-R1-14B-V0.4 using supervised fine-tuning on concise reasoning chains. They also developed V0.6 through iterative curriculum learning. A separate variant, V0.4-RS, incorporated reinforcement learning. The final model emerged from linearly merging the weights of V0.6 and V0.4. This approach combined the strengths of different training methodologies. The team also released Fathom-R1-14B-RS, which merged V0.6 with the reinforcement learning variant, costing 967 US dollars to produce.

Inside the model

The training strategy prioritised efficiency and specific data curation. For the second stage, the team used a dataset of 9,500 difficult problems with low solve rates. They extracted the shortest valid reasoning chains for these questions to stabilise performance within the 16,000-token limit. The final model resulted from merging two checkpoints: one trained with iterative curriculum supervised fine-tuning and another with shortest-chain fine-tuning. This linear merging technique allowed the combination of distinct learning paths. The total post-training cost amounted to 499 US dollars, utilising 176 hours on H100 graphics processing units.

What it can do

On the AIME 2025 mathematics exam, the model achieved a pass@1 accuracy of 52.71 per cent. With consistency checking over 64 samples, this score rose to 76.7 per cent. It scored 35.26 per cent pass@1 on the HMMT 2025 benchmark. These results surpass earlier open-source models like LightR1-14B and proprietary systems such as o1-mini and o3-mini-low. Performance closely rivals o4-mini-low in consistency metrics. The model also demonstrated generalisation beyond mathematics, scoring 59.46 per cent pass@1 on the GPQA-Diamond science benchmark, despite no non-math questions in its training data.

How to use it

Fractal AI Research has open-sourced the model weights, training recipes, and datasets. The code and evaluation frameworks are available on GitHub and Hugging Face. Users can download the checkpoints directly or interact with the model via hosted spaces. The open licence allows the community to inspect the training methods and build upon the work. This accessibility supports further research into cost-effective reasoning models. The low barrier to entry enables developers with limited resources to experiment with advanced post-training techniques.

Why it matters

The project demonstrates that high-quality reasoning performance does not necessarily require expensive infrastructure. By achieving competitive results with a modest budget, it challenges the assumption that only large laboratories can produce top-tier models. The release of the training recipe encourages reproducibility and innovation in the open-source community. It provides a template for efficient post-training on specialised tasks. The success of model merging in this context offers a practical pathway for combining different optimisation strategies without retraining from scratch.

Questions people ask about Fathom-R1-14B

How much did it cost to train Fathom-R1-14B?

The total post-training cost for the final model was 499 US dollars. This covered 176 hours of compute time on H100 graphics processing units. The low cost was achieved by using targeted supervised fine-tuning and model merging rather than extensive reinforcement learning from scratch.

What is the context window size of this model?

Fathom-R1-14B operates with a context window of 16,000 tokens. This constraint was intentional to reflect practical deployment limits and ensure the model could stabilise its reasoning chains within a manageable sequence length during the fine-tuning process.

How does it compare to closed-source models?

On mathematical benchmarks like AIME 2025 and HMMT 2025, it outperforms OpenAI's o1-mini and o3-mini-low models. Its consistency scores closely match those of the newer o4-mini-low model, demonstrating competitive reasoning abilities despite being open-source and significantly smaller.

Can it solve problems outside of mathematics?

Yes, the model shows generalisation to other domains. Although trained exclusively on mathematical questions, it achieved a 59.46 per cent pass@1 score on the GPQA-Diamond benchmark, which tests scientific knowledge. This suggests the reasoning skills learned transfer to diverse subject areas.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Fathom-R1-14B yet, so the facts on this page come from the sources listed above: huggingface.co/FractalAIResearch/Fathom-R1-14B. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.