---
license: other
license_name: ztech-license
license_link: LICENSE
language:
- en
pipeline_tag: text-generation
---

# OmniMath-2B - GGUF Quantized

**This is the official GGUF repository for the OmniMath-2B model.**  
All quantizations provided here are created and maintained by the original model author.

**Original model**: [ZirTech/OmniMath-2B](https://huggingface.co/ZirTech/OmniMath-2B)

---

## Quantizations & File Sizes

| Quantization | File Size | Description |
|--------------|-----------|-------------|
| **Q2_K**     | 969 MB    | 2-bit, smallest, lowest quality |
| **Q3_K_S**   | 1.02 GB   | 3-bit, small |
| **Q3_K_M**   | 1.10 GB   | 3-bit, medium |
| **Q3_K_L**   | 1.16 GB   | 3-bit, large |
| **IQ4_XS**   | 1.20 GB   | 4-bit integer with improved accuracy |
| **Q4_K_S**   | 1.21 GB   | 4-bit, small |
| **Q4_K_M**   | 1.27 GB   | 4-bit, medium (good balance) |
| **Q5_K_S**   | 1.37 GB   | 5-bit, small |
| **Q5_K_M**   | 1.41 GB   | 5-bit, medium |
| **Q6_K**     | 1.56 GB   | 6-bit, high quality |
| **Q8_0**     | 2.01 GB   | 8-bit, near‑original quality |
| **F16**      | 3.78 GB   | 16-bit float (original weights) |

---

# Recommended Quantization

Since **OmniMath-2B** is specialized for mathematical reasoning, **accuracy is paramount**. Lower bit quantizations (Q2, Q3, Q4) may degrade performance on complex problems.

| Recommendation | Quantization | Size   | Notes |
|----------------|--------------|--------|-------|
| **Best for math** (minimal quality loss) | **Q8_0** or **F16** | 2.01 GB / 3.78 GB | Use if you have enough RAM/VRAM. |
| **Good trade‑off** (recommended) | **Q6_K** or **Q5_K_M** | 1.56 GB / 1.41 GB | Still high accuracy, much smaller than F16. |
| **Minimum acceptable** (for tight memory) | **Q4_K_M** | 1.27 GB | May lose some precision; test before using in production. |
| **Not recommended** | Q2_K, Q3_K_*, IQ4_XS | < 1.2 GB | Likely to degrade mathematical reasoning. |

> **Tip**: For the best results, use **Q8_0** or **Q6_K**. If you need to save space, **Q5_K_M** is the lowest we recommend for math.

---

# License

This repository and the quantized files are released under the ztech-license.
Please refer to the original model repository for the full license text.

The GGUF format quantizations are provided by the original author. No third-party ownership is claimed.

---

<div align="center">

**Built by [Zirt Tech](https://huggingface.co/ZirTech) ❤️**

</div>