Qwen3.5-9B - GGUF High-Quality Quantizations
This repository provides GGUF quantized versions of the Qwen/Qwen3.5-9B model, optimized for local execution using llama.cpp and compatible ecosystems.
📌 Version Notes
All quantizations were generated from the official FP16 weights.
- Target: Efficient execution on consumer hardware, mobile/edge devices, and systems with limited memory.
- Performance: The output quality (reasoning, coherence, and accuracy) is strictly dependent on the base model's parameter scale (9B).
📊 Quantization Table
| File | Method | Bit | Description |
|---|---|---|---|
| fp16.gguf | FP16 | 16-bit | Original Weights. No quantization applied. Maximum fidelity. |
| Q8_0.gguf | Q8_0 | 8-bit | Near-lossless. Practically identical to the original model with lower memory footprint. |
| Q5_K_M.gguf | Q5_K_M | 5-bit | High Precision. Minimizes quantization error for critical tasks. |
| Q4_K_M.gguf | Q4_K_M | 4-bit | Recommended. Best balance between speed and performance. |
| Q4_K_S.gguf | Q4_K_S | 4-bit | Fast/Small. Optimized for maximum throughput and low RAM usage. |
🛠️ Technical Details
- Quantization Date: 2026-03-07
- Tool used:
llama-quantize(llama.cpp) - Method: K-Quantization (optimized for AVX2/AVX-512 and modern GPU architectures).
🚀 How to Use
Start a local OpenAI-compatible server with a web UI:
llama.cpp (CLI) using model from HuggingFace
./llama-cli -hf daniloreddy/Qwen3.5-9B_GGUF:Q4_K_M -p "User: Hello! Assistant:" -n 512 --temp 0.7
llama.cpp (CLI) using downloaded model
./llama-cli -m path/to/Qwen3.5-9B_Q4_K_M.gguf -p "User: Hello! Assistant:" -n 512 --temp 0.7
llama.cpp (SERVER) using model from HuggingFace
./llama-server -hf daniloreddy/Qwen3.5-9B_GGUF:Q4_K_M --port 8080 -c 4096
llama.cpp (SERVER) using downloaded model
./llama-server -m /path/to/Qwen3.5-9B_Q4_K_M.gguf --port 8080 -c 4096