Xinghe1-9B-GGUF

Xinghe1-9B-GGUF (杏核1代-9B GGUF)

Xinghe1-9B-GGUF (杏核) contains the quantized GGUF versions of the Xinghe1-9B model, which is fine-tuned from the Qwen3.5-9B-Instruct architecture for the formalization, computational derivation, and clinical reasoning of Huangdi Neijing.

This repository provides multiple quantization formats compiled using llama-quantize for efficient local deployment and inference.

Released Quantizations

• Xinghe1-9B-BF16.gguf (16.69 GB) — Native best-quality Bfloat16 precision, recommended for GPU setups.
• Xinghe1-9B-Q8_0.gguf (8.87 GB) — 8-bit high-quality quantization, minimal quality loss.
• Xinghe1-9B-Q6_K.gguf (6.85 GB) — 6-bit balanced quantization, recommended for most setups.
• Xinghe1-9B-Q4_K_M.gguf (5.24 GB) — 4-bit fast and lightweight quantization.

Note: All GGUF models in this repository have been fixed to exclude the Qwen3.5 MTP (Multi-Token Prediction) layers, avoiding metadata layer count conflicts. They can be directly loaded into LM Studio without errors.

Deployment & Usage

1. LM Studio

Search for zsyjsld/Xinghe1-9B-GGUF directly in the LM Studio search bar, select your preferred quantization version, and load the model.

2. Ollama

Run the 4-bit version using the following command:

ollama run hf.co/zsyjsld/Xinghe1-9B-GGUF:Q4_K_M

Model Details

• Developed by: zsyjsld
• Base Model: Qwen/Qwen3.5-9B-Instruct
• Fine-tuning Method: QLoRA SFT on V3 Double-Purity dataset.
• Language(s): English & LaTeX (Internal reasoning formulas), English (Output explanations)
• License: Apache 2.0 (for code/dataset) & Qwen License Agreement (for model weights)

Technical Architecture

Xinghe1-9B is trained to strictly bind clinical concepts and perform dynamical derivations inside the <think> tag, and output natural, clean clinical TCM analyses outside the <think> tag.

The internal derivation is based on the unified meta-framework:

$$\dot{x} = F(x) + G(x, \mu(t))$$

$$y = h(x) + \epsilon$$