# JT-LM/JT-Math-8B-Thinking-GGUF

This repository contains GGUF format model files converted from [JT-LM/JT-Math-8B-Thinking](https://huggingface.co/JT-LM/JT-Math-8B-Thinking), optimized for `llama.cpp` and other GGUF-compatible inference clients (such as LM Studio, Ollama, AnythingLLM, etc.).

## Model Overview
**JT-Math-8B-Thinking** is an 8-billion parameter open-source Large Language Model designed specifically for **advanced mathematical reasoning** and **complex problem-solving**. Fine-tuned on high-quality bilingual (Chinese and English) datasets, the model features strong long-context processing capabilities and powerful Chain-of-Thought (CoT) reasoning.

- **Key Features**:
  - **Long Context Support**: Natively supports up to a **32,768 (32K)** context window.
  - **Deep Reasoning**: Optimized via multi-stage Reinforcement Learning (RL) and curriculum learning, making it exceptionally good at generating deep reasoning paths to solve competition-level math problems.
  - **Bilingual Optimization**: Delivers top-tier mathematical derivation performance in both Chinese and English environments.

---

## File List & Quantization Options

This repository offers two high-precision versions, ideal for scenarios that demand ultimate reasoning quality and have sufficient hardware resources:

| File Name | Type | File Size | Recommended RAM/VRAM | Description |
| :--- | :--- | :---: | :---: | :--- |
| `JT-Math-8B-Thinking-Q8_0.gguf` | Q8_0 Quantization | ~8.5 GB | >= 12 GB | **Recommended Choice.** Almost lossless 8-bit quantization that perfectly balances inference speed and model performance. Suitable for most modern CPUs and GPUs. |
| `JT-Math-8B-Thinking-F16.gguf`  | F16 Native | ~16.1 GB | >= 24 GB | **Lossless Version.** Retains the original Float16 precision. Ideal for resource-rich environments (e.g., 24GB VRAM GPUs) where any quantization loss is unacceptable. |