Qwen3-Coder 4bit TQ Released

textclf has released Qwen3-Coder-Next-TQ-4bit, a 4bit quantized version of Qwen3-Coder-Next. Unlike traditional calibration-based methods, this model applies the TQ (Training-free Quantization) technique, which requires no calibration process.

Features of TQ Quantization

• Calibration-free approach: By not relying on calibration data, it avoids bias toward specific tasks and provides better generalization performance on downstream tasks.
• Performance: Maintains performance levels equivalent to existing quantization methods on the KL Divergence (KLD) metric.
• Size: The model size is 41.7GB.

Deployment and Execution

It can be run with vLLM via Docker, enabling TQ quantization with the --quantization tq_quant option. The tool-call parser used is qwen3_coder.