Fathom-R1-14B GGUF Models
Model Generation Details
This model was generated using llama.cpp at commit ea1431b0.
Choosing the Right Model Format
Selecting the correct model format depends on your hardware capabilities and memory constraints.
BF16 (Brain Float 16) – Use if BF16 acceleration is available
- A 16-bit floating-point format designed for faster computation while retaining good precision.
- Provides similar dynamic range as FP32 but with lower memory usage.
- Recommended if your hardware supports BF16 acceleration (check your device's specs).
- Ideal for high-performance inference with reduced memory footprint compared to FP32.
📌 Use BF16 if:
✔ Your hardware has native BF16 support (e.g., newer GPUs, TPUs).
✔ You want higher precision while saving memory.
✔ You plan to requantize the model into another format.
📌 Avoid BF16 if:
❌ Your hardware does not support BF16 (it may fall back to FP32 and run slower).
❌ You need compatibility with older devices that lack BF16 optimization.
F16 (Float 16) – More widely supported than BF16
- A 16-bit floating-point high precision but with less of range of values than BF16.
- Works on most devices with FP16 acceleration support (including many GPUs and some CPUs).
- Slightly lower numerical precision than BF16 but generally sufficient for inference.
📌 Use F16 if:
✔ Your hardware supports FP16 but not BF16.
✔ You need a balance between speed, memory usage, and accuracy.
✔ You are running on a GPU or another device optimized for FP16 computations.
📌 Avoid F16 if:
❌ Your device lacks native FP16 support (it may run slower than expected).
❌ You have memory limitations.
Quantized Models (Q4_K, Q6_K, Q8, etc.) – For CPU & Low-VRAM Inference
Quantization reduces model size and memory usage while maintaining as much accuracy as possible.
- Lower-bit models (Q4_K) → Best for minimal memory usage, may have lower precision.
- Higher-bit models (Q6_K, Q8_0) → Better accuracy, requires more memory.
📌 Use Quantized Models if:
✔ You are running inference on a CPU and need an optimized model.
✔ Your device has low VRAM and cannot load full-precision models.
✔ You want to reduce memory footprint while keeping reasonable accuracy.
📌 Avoid Quantized Models if:
❌ You need maximum accuracy (full-precision models are better for this).
❌ Your hardware has enough VRAM for higher-precision formats (BF16/F16).
Very Low-Bit Quantization (IQ3_XS, IQ3_S, IQ3_M, Q4_K, Q4_0)
These models are optimized for extreme memory efficiency, making them ideal for low-power devices or large-scale deployments where memory is a critical constraint.
IQ3_XS: Ultra-low-bit quantization (3-bit) with extreme memory efficiency.
- Use case: Best for ultra-low-memory devices where even Q4_K is too large.
- Trade-off: Lower accuracy compared to higher-bit quantizations.
IQ3_S: Small block size for maximum memory efficiency.
- Use case: Best for low-memory devices where IQ3_XS is too aggressive.
IQ3_M: Medium block size for better accuracy than IQ3_S.
- Use case: Suitable for low-memory devices where IQ3_S is too limiting.
Q4_K: 4-bit quantization with block-wise optimization for better accuracy.
- Use case: Best for low-memory devices where Q6_K is too large.
Q4_0: Pure 4-bit quantization, optimized for ARM devices.
- Use case: Best for ARM-based devices or low-memory environments.
Summary Table: Model Format Selection
| Model Format | Precision | Memory Usage | Device Requirements | Best Use Case |
|---|---|---|---|---|
| BF16 | Highest | High | BF16-supported GPU/CPUs | High-speed inference with reduced memory |
| F16 | High | High | FP16-supported devices | GPU inference when BF16 isn't available |
| Q4_K | Medium Low | Low | CPU or Low-VRAM devices | Best for memory-constrained environments |
| Q6_K | Medium | Moderate | CPU with more memory | Better accuracy while still being quantized |
| Q8_0 | High | Moderate | CPU or GPU with enough VRAM | Best accuracy among quantized models |
| IQ3_XS | Very Low | Very Low | Ultra-low-memory devices | Extreme memory efficiency and low accuracy |
| Q4_0 | Low | Low | ARM or low-memory devices | llama.cpp can optimize for ARM devices |
Included Files & Details
Fathom-R1-14B-bf16.gguf
- Model weights preserved in BF16.
- Use this if you want to requantize the model into a different format.
- Best if your device supports BF16 acceleration.
Fathom-R1-14B-f16.gguf
- Model weights stored in F16.
- Use if your device supports FP16, especially if BF16 is not available.
Fathom-R1-14B-bf16-q8_0.gguf
- Output & embeddings remain in BF16.
- All other layers quantized to Q8_0.
- Use if your device supports BF16 and you want a quantized version.
Fathom-R1-14B-f16-q8_0.gguf
- Output & embeddings remain in F16.
- All other layers quantized to Q8_0.
Fathom-R1-14B-q4_k.gguf
- Output & embeddings quantized to Q8_0.
- All other layers quantized to Q4_K.
- Good for CPU inference with limited memory.
Fathom-R1-14B-q4_k_s.gguf
- Smallest Q4_K variant, using less memory at the cost of accuracy.
- Best for very low-memory setups.
Fathom-R1-14B-q6_k.gguf
- Output & embeddings quantized to Q8_0.
- All other layers quantized to Q6_K .
Fathom-R1-14B-q8_0.gguf
- Fully Q8 quantized model for better accuracy.
- Requires more memory but offers higher precision.
Fathom-R1-14B-iq3_xs.gguf
- IQ3_XS quantization, optimized for extreme memory efficiency.
- Best for ultra-low-memory devices.
Fathom-R1-14B-iq3_m.gguf
- IQ3_M quantization, offering a medium block size for better accuracy.
- Suitable for low-memory devices.
Fathom-R1-14B-q4_0.gguf
- Pure Q4_0 quantization, optimized for ARM devices.
- Best for low-memory environments.
- Prefer IQ4_NL for better accuracy.
🚀 If you find these models useful
❤ Please click "Like" if you find this useful!
Help me test my AI-Powered Network Monitor Assistant with quantum-ready security checks:
👉 Free Network Monitor
💬 How to test:
Choose an AI assistant type:
TurboLLM(GPT-4o-mini)HugLLM(Hugginface Open-source)TestLLM(Experimental CPU-only)
What I’m Testing
I’m pushing the limits of small open-source models for AI network monitoring, specifically:
- Function calling against live network services
- How small can a model go while still handling:
- Automated Nmap scans
- Quantum-readiness checks
- Network Monitoring tasks
🟡 TestLLM – Current experimental model (llama.cpp on 2 CPU threads):
- ✅ Zero-configuration setup
- ⏳ 30s load time (slow inference but no API costs)
- 🔧 Help wanted! If you’re into edge-device AI, let’s collaborate!
Other Assistants
🟢 TurboLLM – Uses gpt-4o-mini for:
- Create custom cmd processors to run .net code on Free Network Monitor Agents
- Real-time network diagnostics and monitoring
- Security Audits
- Penetration testing (Nmap/Metasploit)
- 🔑 Get more tokens by logging in or downloading our Free Network Monitor Agent with integrated AI Assistant
🔵 HugLLM – Latest Open-source models:
- 🌐 Runs on Hugging Face Inference API
💡 Example commands to you could test:
"Give me info on my websites SSL certificate""Check if my server is using quantum safe encyption for communication""Run a comprehensive security audit on my server"- '"Create a cmd processor to .. (what ever you want)" Note you need to install a Free Network Monitor Agent to run the .net code from. This is a very flexible and powerful feature. Use with caution!
🧮 Fathom-R1-14B: $499 Training Recipe for Unlocking Math Reasoning at o4-mini level using R1-distilled-14B model under 16K context
Overview
Reasoning models often require high post-training budgets and extremely long reasoning chains(think 32k/64k) for maximising performance. Can we improve these models even if both these parameters are capped?
To this end, we first introduce: Fathom-R1-14B, a 14-billion-parameter reasoning language model derived from Deepseek-R1-Distilled-Qwen-14B, post-trained at an affordable cost of only $499, and achieving SOTA mathematical reasoning performance within a 16K context window. On the latest olympiad level exams: AIME-25 and HMMT-25, our model not only surpasses o3-mini-low, o1-mini and LightR1-14B(16k) at pass@1 scores (averaged over 64 runs) but also delivers performance rivaling closed-source o4-mini (low) w.r.t cons@64 — all while staying within a 16K context window. It achieves 52.71% Pass@1 accuracy on AIME2025 and 35.26% Pass@1 accuracy on HMMT25 (+7.2% and +5.2% improvement over the base model respectively). When provided with additional test-time compute in the form of cons@64, it achieves an impressive 76.7% accuracy on AIME2025 and 56.7% accuracy on HMMT25 (+13.4% and +6.7% improvement over the base model respectively). We perform supervised fine-tuning (SFT) on carefully curated datasets using a specific training approach, followed by model merging, achieving this performance at a total cost of just $499!
We also introduce Fathom-R1-14B-RS, another model achieving performance comparable to our first, at a total post-training cost of just $967. It leverages post-training techniques—including reinforcement learning and supervised fine-tuning—in a multi-stage, cost-effective manner, followed by model merging.
We are open-sourcing our models, post-training recipes and datasets which we believe will help the community to progress further in the reasoning domain.
🧪 Motivation
Thinking longer during inference time has shown to unlock superior reasoning abilities and expert level performance on challenging queries and tasks. Since the open-source release of DeepSeek R1 series models, multiple open-source efforts [s1, LIMO, Light-R1] have focused on reproducing the results (easpecially at <=32B scale) either via distillation or RL based fine-tuning on top of non-reasoning models. Though in most cases, these efforts at best, could come close to the performance R1 series models but are unable to surpass them. In parallel, certain recent methods [DeepScaleR, DeepCoder, Light-R1] started with the existing reasoning models and have managed to extend the performance of these models. However, the training runs for these methods are often costly and they rely on longer sequence lengths for higher accuracy.
Given the latest findings [Proof or Bluff ?, Reasoning models don't always say what they think] that raises the question on the correctness of the intermediate steps of the long COT in reasoning models, its important from interpretability, reliability and safety pov to ensure the reasoning chains are not inefficiently long. Hence, in this study, we work towards unlocking performance improvement of the reasoning models without training at very high (24k/32k) sequence length and restricting it to 16k context. We believe, while extremely long reasoning chains are still necessary for really challenging tasks, its also important to maximize the performance at lower context first before we proceed towards extending reasoning chains.
Training Dataset
We begin by curating a high-quality mathematical corpus from the following open-source datasets:
- Open-R1 - default subset
- Numina – Olympiads & AOPS_forum (word problems, float type answers)
- After rigorous deduplication and decontamination, we consolidated approximately ~100K unique problems, forming the initial corpus for all subsequent trainings.
🏗️ Post-Training Strategies
Training Recipe for Fathom-R1-14B-v0.6
SFT on difficult questions and their reasoning chains has been shown to be effective for improving reasoning ability. For this checkpoint, we build on top of this. This training stage focuses on improving the model’s performance on mathematical problems covering a spectrum of hard diffuculty level through a iterative curriculum learning strategy at max 16k sequence length. Curriculum learning (CL) is a well-established technique for training LLMs, where the model is progressively exposed to more difficult tasks. The idea is to gradually scaffold more complex reasoning, thereby enhancing generalization and reducing overfitting. However, in our case we perform this in an iterative manner, which essentially means we do multiple iterations of CL.
For the dataset preparation, we begin by annotating each question’s difficulty using OpenAI's o3mini model. We retain only those questions rated above average (in relative sense) and further filter them to include only those having solve rates between certain range (0.2 < pass_rate < 0.7). This yields the Iterative Curriculum Learning dataset comprising of 5K examples.
Total H100 GPU Hours: 48
Cost: $136
Training Recipe for Fathom-R1-14B-v0.4-RS
The core strategy used behind creating this checkpoint is a two-stage pipeline: First, leverage GRPO to improve reasoning of Deepseek-R1-Distilled-Qwen-14B at a lower sequence length, 6k, on a carefully curated dataset to ensure rapid improvement with minimal training steps. Second, we perform SFT, at max 16k tokens sequence length, on a carefully curated dataset of questions ( hard to very hard difficulty spectrum) and the corresponding shortest possible reasoning solution for each question.
- First Stage (Leveraging RL for effecient test-time thinking): We start with curating a seed dataset which ensures the policy receives minumium reward while still having room for growth. The dataset comparises of questions having solve rates (at lower sequence length) between a certain range. This forms our RL Compression dataset comprising of 7.7K questions. Staring from DeepSeek-R1-Distill-Qwen-14B as the base model, we train the model using the GRPO algorithm, with a 6k token sequence length limit. We see a consistent increase in performance as the model learns to generate concise responses from the decreasing clip ratio, response length and increasing reward. The obtained model has learnt to generate responses below 6k tokens and outperforms the base model at lower token limits.
- Second Stage (Leveraging SFT to improve reasoning efficiently at higher sequence length): We build upon the RL checkpoint and perform SFT under a 16K context window to encourage more detailed reasoning that would be required for solving more complex problems. For this stage, we strategically curate a dataset consisting of hard problems — specifically, questions with lower solve rates (0 < pass_rate <=0.4). Then, we obtain the shortest possible reasoning chains for these questions forming the SFT Shortest Chains dataset comprising of 9.5K examples. Through supervised fine-tuning on this dataset, the model is able to stablize its reasoning at sequence length upto 16K. The resulting model is named Fathom-R1-14B-v0.4, optimized for concise yet accurate mathematical reasoning.
Total H100 GPU Hours: 293
Cost: $831
Training Recipe for Fathom-R1-14B-v0.4
Given the performance improvement we noticed during the second fine-tuning stage of developing Fathom-R1-14B-v0.4-RS and in attempt to further reduce the cost, we experiment by eliminating RL and directly performing second stage SFT on Deepseek-R1-Distilled-Qwen-14B base model.
Total H100 GPU Hours: 128
Cost: $363
Model Merging
Given v0.6 and v0.4 models have been developed by following different training methodologies, we perform linear merging to combine the strengths to obtain final 2 checkpoints.
- Fathom-R1-14B: Obtained via merging Fathom-R1-14B-V0.6 (Iterative Curriculum SFT) and Fathom-R1-14B-V0.4 (SFT-Shortest-Chains)
- Fathom-R1-14B-RS: Obtained via merging Fathom-R1-14B-V0.6 (Iterative Curriculum SFT) and Fathom-R1-14B-V0.4 (RL-compression + SFT-Shortest-Chains)
💰 Post-Training Cost
We developed Fathom-R1-14B models using a focused, resource-efficient strategy that balances performance with compute budget. Below is the GPU time utilized and the cost incurred
| Model Weights | GPU Hours (H100) | Cost(USD) |
|---|---|---|
| Fathom-R1-14B-V0.4-RS | 293 | 831 |
| Fathom-R1-14B-V0.4 | 128 | 363 |
| Fathom-R1-14B-V0.6 | 48 | 136 |
| Fathom-R1-14B-RS | 341 | 967 |
| Fathom-R1-14B | 176 | 499 |
So, the final Fathom-R1-14B took just 499$ to be trained overall! This low training cost highlights the efficiency of our method — enabling high-level mathematical reasoning comparable to o4-mini in $499 , all within a 16k sequence length budget.
📊 Evaluation
We evaluate Fathom‑R1-14B using the same metrics and sampling configuration introduced in the DeepSeek‑R1 paper, namely pass@1 and cons@64. However, our evaluation is conducted under a reduced output budget of 16,384 tokens, compared to DeepSeek‑R1’s 32,768 tokens, to better reflect practical deployment constraints.
- pass@1: Pass@1 is computed as the average correctness over k sampled solution chains per problem (in our experiments we keep k=64).
- cons@64: Assesses consistency by sampling 64 reasoning chains per question and computing the majority vote accuracy.
Evaluation Configuration:
- Temperature: 0.6
- top_p: 0.95
- Number of sampled chains: 64
- Context: 16,384 tokens
This setup allows us to benchmark Fathom-R1-14B’s reasoning performance and stability under realistic memory and inference budgets, while maintaining compatibility with the DeepSeek‑R1 evaluation protocol.
We utilize the evaluation framework provided by the LIMO repository to run inference and compute metrics.
For detailed instructions and implementation details, please refer to eval/README.md.
Results
We evaluate and compare Fathom‑R1-14B with several baseline models across 3 challenging benchmarks: AIME25, HMMT25, and GPQA. For each, we report pass@1 and cons@64, following the same evaluation configuration.
| Model | AIME25 | HMMT25 | ||
|---|---|---|---|---|
| pass@1 | cons@64 | pass@1 | cons@64 | |
| Closed-Source Models | ||||
| o1‑mini | 50.71 | 63.33 | 35.15 | 46.67 |
| o3‑mini‑low | 42.60 | 53.33 | 26.61 | 33.33 |
| o3‑mini‑medium | 72.24 | 83.33 | 49.21 | 60.00 |
| o4-mini-low | 60.20 | 76.67 | 39.11 | 53.33 |
| o1‑preview | 33.33 | 36.67 | 17.78 | 20.00 |
| gpt‑4.5‑preview | 34.44 | 40.00 | 16.67 | 20.00 |
| Open-Source Models | ||||
| DeepSeek-R1-Distill-Qwen-14B | 45.50 | 63.33 | 30.00 | 50.00 |
| DeepSeek-R1-Distill-Qwen-32B | 49.64 | 73.33 | 33.02 | 53.33 |
| DeepSeekR1‑670B | 61.25 | 83.33 | 42.19 | 56.67 |
| LightR1‑14B | 51.15 | 76.67 | 33.75 | 50.00 |
| Fathom‑R1-14B-V0.4-RS | 50.94 | 73.33 | 33.70 | 40.00 |
| Fathom‑R1-14B-V0.4 | 50.94 | 70.00 | 34.53 | 56.67 |
| Fathom‑R1-14B-V0.6 | 50.63 | 76.67 | 32.19 | 50.00 |
| Fathom‑R1-14B-RS | 52.03 | 76.67 | 35.00 | 53.33 |
| Fathom‑R1-14B | 52.71 | 76.67 | 35.26 | 56.67 |
Fathom‑R1-14B demonstrates highly competitive performance across all datasets, improving over the original R1-distilled models while closely matching or surpassing other strong baselines in several settings. On both AIME 25 and HMMT 25, our model shows the highest pass@1 as well as cons@64 scores among all the open-source models (including the bigger R1-Distilled-32B model), with R1-670B being the only exception.
In fact, we observe that Fathom-R1-14B is superior to the first two generations of OpenAI's mini-reasoning models, including o1-mini and o3-mini-low- and it's performance closely matches that of newly released o4-mini-low (self-consistency decoding).
🌍 Generalization Beyond Math: GPQA-Diamond
Notably, we also observe out-of-domain improvement in GPQA-Diamond, even though there wasn't a single instance of non-math questions in our training data. This indicates that our training methodology mentioned above and training on math wuestions facilitates generalization across diverse domains, a finding similar to what LightR1-14B & LIMO had observed.
✅ GPQA Benchmark Comparison (16k)
| Model | pass@1 | cons@64 |
|---|---|---|
| DeepSeek-R1-Distill-Qwen-14B | 54.19 | 64.14 |
| LightR1‑14B | 56.94 | 65.15 |
| Fathom‑R1-14B-RS | 59.13 | 66.16 |
| Fathom‑R1-14B | 59.46 | 66.16 |
✂️ Ablation Study on Token Efficiency
To assess reasoning token efficiency, we compare the average response token count, under 16k context length, across AIME25, and HMMT25. On AIME25, Fathom‑R1-14B-RS uses 10% fewer response tokens than LightR1-14B despite having higher pass@1. HMMT25 questions are relatively tougher compared to AIME'25 and tougher questions usually require more thinking tokens. On HMMT25, Fathom‑R1-14B-RS uses 4.5% fewer response tokens than LightR1-14B despite having higher pass@1.
Average Response Length (Tokens)
| Model | AIME25 | HMMT25 |
|---|---|---|
| LightR1-14B | 11330 | 12680 |
| DeepSeek-R1-Distill-Qwen-14B | 10878 | 12263 |
| Fathom‑R1-14B-V0.4 | 10570 | 11950 |
| Fathom‑R1-14B | 10956 | 12125 |
| Fathom‑R1-14B-RS | 10083 | 12100 |
Data Decontimination
Both benchmarks used (AIME 25 and HMMT 25) were released a few weeks after the release of the base model, ensuring no contamination occurred during the model's pre-training. The dataset corpora (Numina-Math 1.5 & OpenR1-Math) were released around the same time as these exams, with a cutoff date no later than 2024. Additionally, we conduct checks to verify there is no contamination in the training data.
Release Assets
- Training Recipe Blog: 🤗 $499 training recipe for creating Fathom-R1-14B
- Final Merged Models: 🤗 Fathom-R1-14B, 🤗 Fathom-R1-14B-RS
- Intermediate Models: 🤗 Fathom-R1-14B-V0.6, 🤗 Fathom-R1-14B-V0.4, 🤗 Fathom-R1-14B-V0.4-RS
- Fathom-R1-14B Datasets: 🤗 V0.6-Iterative-Curriculum-Learning, 🤗 V0.4-SFT-Shortest-Chains, 🤗 V0.4-RL-Compression
📜 License
This repository and all the release assets are available under the MIT License, underscoring our dedication to open and inclusive AI innovation. By freely sharing our work, we aim to democratize AI technology, empowering researchers, developers, and enthusiasts everywhere to use, adapt, and expand upon it without limitation. This open and permissive approach promotes global collaboration, accelerates innovation, and enriches the AI community as a whole.
Acknowledgments
We would like to acknowledge the following works for enabling our project:
📖 Citation
@misc{fathom14b2025,
title={Fathom-R1: $499 Training Recipe for Unlocking Math Reasoning at o4-mini level with just 14B parameters under 16K context},
author={Kunal Singh and Pradeep Moturi and Ankan Biswas and Siva Gollapalli and Sayandeep Bhowmick},
howpublished={\url{https://huggingface.co/FractalAIResearch/Fathom-R1-14B}},
note={Hugging Face},
year={2025}
}
About Project Ramanujan
We initiated Project Ramanujan approximately one year ago, aiming to unlock intelligence and enhance AI agents by pushing the boundaries of advanced reasoning. Our key accomplishments include:
- ICLR'25 & NeurIPS'24-MATH-AI: SBSC: Step-By-Step Coding for Improving Mathematical Olympiad Performance
- Winners of HackerCupAI@NeurIPS'24 & ICLR'25-VerifAI: Stress Testing Based Self-Consistency For Olympiad Programming
- CVPR'25-MULA: TRISHUL: Towards Region Identification and Screen Hierarchy Understanding for Large VLM based GUI Agents )
- Silver Medal in AIMO'24
- Downloads last month
- 985
1-bit
2-bit
3-bit
4-bit
5-bit
6-bit
8-bit
16-bit
Model tree for Mungert/Fathom-R1-14B-GGUF
Base model
deepseek-ai/DeepSeek-R1-Distill-Qwen-14B