btbtyler09/Qwen3-Coder-Next-GPTQ-4bit

Qwen3-Coder-Next GPTQ 4-bit

GPTQ 4-bit quantization of Qwen/Qwen3-Coder-Next, an 80B-parameter Mixture-of-Experts (MoE) coding model with 3B activated parameters per token.

Model Overview

• Architecture: Qwen3NextForCausalLM (hybrid linear + full attention with DeltaNet)
• Total parameters: ~80B
• Activated parameters: ~3B per token (10 of 512 experts selected per token)
• Layers: 48 (36 linear attention + 12 full attention, repeating 3:1 pattern)
• Experts: 512 per layer + 1 shared expert per layer
• Context length: 262,144 tokens
• Supports: Tool calling, code generation, general chat

Quantization Details

All 73,728 MoE expert modules (512 experts x 3 projections x 48 layers) are quantized to INT4 using GPTQ. Non-expert modules remain at FP16 for quality preservation.

Component	Precision	Notes
MoE experts (gate_proj, up_proj, down_proj)	INT4 (GPTQ)	73,728 modules quantized
Attention (q_proj, k_proj, v_proj, o_proj)	FP16	Full precision
Linear attention (in_proj_qkvz, out_proj, in_proj_ba)	FP16	Full precision
Shared experts	FP16	Full precision
Embeddings, LM head, norms	FP16	Full precision

GPTQ configuration:

• Bits: 4
• Group size: 32
• Symmetric: Yes
• desc_act: No
• true_sequential: Yes
• Failsafe: RTN for poorly-calibrated rare experts (7,650 of 73,728 modules, ~10.4%)

Calibration

• Dataset: Mixed - evol-codealpaca-v1 (code) + C4 (general text)
• Samples: 2,048 with context length binning (uniform distribution across 256-2048 token bins)
• Quantizer: GPTQModel v5.7.0

See quantize.py for the full quantization script.

Model Size

Version	Size	Compression
BF16 (original)	~160 GB	-
GPTQ 4-bit	47 GB	3.4x

Perplexity

Evaluated on wikitext-2-raw-v1 (test set), seq_len=2048, stride=512:

Model	Perplexity	Degradation
BF16 (original)	6.9401	-
GPTQ 4-bit	6.9956	+0.8%

Usage

vLLM (Recommended)

vllm serve btbtyler09/Qwen3-Coder-Next-GPTQ-4bit \
  --tensor-parallel-size 4 \
  --trust-remote-code \
  --quantization gptq \
  --max-model-len 32768

Tool Calling

This model supports tool calling via the Qwen3-Coder chat template. The quantized model includes:

• chat_template.jinja - Chat template with tool support
• qwen3coder_tool_parser_vllm.py - vLLM tool parser plugin
• qwen3_coder_detector_sgl.py - SGLang tool detector

For vLLM tool calling:

vllm serve btbtyler09/Qwen3-Coder-Next-GPTQ-4bit \
  --tensor-parallel-size 4 \
  --trust-remote-code \
  --dtype float16 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3_coder

Credits

• Base Model: Qwen - Qwen3-Coder-Next
• Quantization: GPTQ via GPTQModel v5.7.0
• Quantized by: btbtyler09

License

This model inherits the Apache 2.0 license from the base model.