TextCLF Releases Calibration-Free TQ Quantization Method and Open-Source Quant Factory

TextCLF has introduced TQ, a calibration-free quantization method designed to allow immediate quantization of new models without the need for calibration datasets. The approach aims to improve generalization compared to traditional calibration-based techniques while maintaining comparable performance metrics.

Performance and Methodology

The method was tested on Qwen 3.8 27B, where the 4-bit TQ quantization achieved a mean KLD of 0.0282 and a top-1 accuracy of 92.4%. These results demonstrate that TQ performs closely to calibration-based methods despite eliminating the data collection step required by those approaches.

Open-Source Tools and Availability

The underlying code has been released as Quant Factory on GitHub, enabling developers to quantize and run open-source models locally. Key features include:

• Current Support: 4-bit quantization is currently available.
• Future Roadmap: Support for 2-bit and 3-bit quantization is planned for upcoming releases.
• Deployment: Pre-quantized models are available on Hugging Face, and a Docker image (textclf/tq-quant:4bit-main) is provided for easy deployment via vLLM.

Users can run models such as textclf/Qwen3.8-27B-TQ-4bit using the provided Docker command, which integrates directly with vLLM serving infrastructure.