GLM-4.5-Air-AWQ-FP16Mix

Base model: zai-org/GLM-4.5-Air

【vLLM Single Node with 8 GPUs Startup Command】

Note: You must use --enable-expert-parallel to start this model, otherwise the expert tensor TP will not divide evenly. This is required even for 2 GPUs.

CONTEXT_LENGTH=32768

vllm serve \
    QuantTrio/GLM-4.5-Air-AWQ-FP16Mix \
    --served-model-name GLM-4.5-Air-AWQ-FP16Mix \
    --enable-expert-parallel \
    --swap-space 16 \
    --max-num-seqs 512 \
    --max-model-len $CONTEXT_LENGTH \
    --max-seq-len-to-capture $CONTEXT_LENGTH \
    --gpu-memory-utilization 0.9 \
    --tensor-parallel-size 8 \
    --trust-remote-code \
    --disable-log-requests \
    --host 0.0.0.0 \
    --port 8000

【Dependencies】

vllm==0.10.0

【❗❗Temporary Patch for vllm==0.10.0❗❗】

The awq_marlin module in vllm misses checking the modules_to_not_convert parameter when loading AWQ-MoE modules, which causes mixed quantization of MoE to fail or report errors.
Refer to: [Issue #21888]

Before the PR is merged, temporarily replace awq_marlin.py in vllm/model_executor/layers/quantization/awq_marlin.py.

【Model Update Date】

2025-07-30
1. Initial commit

【Model Files】

【Model Download】

from huggingface_hub import snapshot_download
snapshot_download('QuantTrio/GLM-4.5-Air-AWQ-FP16Mix', cache_dir="your_local_path")

【Overview】

GLM-4.5

Model Introduction

The GLM-4.5 model series is a foundation model designed for agents. GLM-4.5 has 355 billion total parameters, of which 32 billion are active. GLM-4.5-Air adopts a more compact design with 106 billion total parameters and 12 billion active parameters. The GLM-4.5 models unify reasoning, encoding, and agent capabilities to meet the complex demands of agent applications.

Both GLM-4.5 and GLM-4.5-Air are hybrid reasoning models that offer two modes: a thinking mode for complex reasoning and tool use, and a non-thinking mode for instant responses.

We have open-sourced the base models, hybrid reasoning models, and FP8 versions of GLM-4.5 and GLM-4.5-Air. They are released under the MIT license and can be used for commercial purposes and secondary development.

In our comprehensive evaluation across 12 industry-standard benchmarks, GLM-4.5 achieved an excellent score of 63.2, ranking 3rd among all proprietary and open-source models. Notably, GLM-4.5-Air maintained strong efficiency while achieving a competitive score of 59.8.

For more detailed evaluation results, demo cases, and technical information, please visit our technical blog. The full technical report will be released soon.

Model code, tool parsers, and inference parsers can be found in the following implementations:

• transformers
• vLLM
• SGLang

Quick Start

Please refer to our GitHub project.