GLM-4.5-FP8

GLM-4.5-FP8#

📚 Paper | 💻 Code | 🌐 Project Page

Image: image

👋 Join our Discord community.

📖 Check out the GLM-4.5 technical blog.

📍 Use GLM-4.5 API services on Z.ai API Platform (Global) or Zhipu AI Open Platform (Mainland China).

👉 One click to GLM-4.5.

Paper Abstract#

We present GLM-4.5, an open-source Mixture-of-Experts (MoE) large language model with 355B total parameters and 32B activated parameters, featuring a hybrid reasoning method that supports both thinking and direct response modes. Through multi-stage training on 23T tokens and comprehensive post-training with expert model iteration and reinforcement learning, GLM-4.5 achieves strong performance across agentic, reasoning, and coding (ARC) tasks, scoring 70.1% on TAU-Bench, 91.0% on AIME 24, and 64.2% on SWE-bench Verified. With much fewer parameters than several competitors, GLM-4.5 ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks. We release both GLM-4.5 (355B parameters) and a compact version, GLM-4.5-Air (106B parameters), to advance research in reasoning and agentic AI systems. Code, models, and more information are available at this https URL .

Model Introduction#

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

Both GLM-4.5 and GLM-4.5-Air are hybrid reasoning models that provide two modes: thinking mode for complex reasoning and tool usage, and non-thinking mode for immediate responses.

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

As demonstrated in our comprehensive evaluation across 12 industry-standard benchmarks, GLM-4.5 achieves exceptional performance with a score of 63.2, in the 3rd place among all the proprietary and open-source models. Notably, GLM-4.5-Air delivers competitive results at 59.8 while maintaining superior efficiency.

Image: bench

For more eval results, show cases, and technical details, please visit our technical blog or refer to the technical report (paper).

The model code, tool parser and reasoning parser can be found in the implementation of transformers, vLLM and SGLang.

Model Downloads#

You can directly experience the model on Hugging Face or ModelScope or download the model by following the links below.

🤖 ModelScope | 355B-A32B | BF16 | | GLM-4.5-Air | 🤗 Hugging Face 🤖 ModelScope | 106B-A12B | BF16 | | GLM-4.5-FP8 | 🤗 Hugging Face 🤖 ModelScope | 355B-A32B | FP8 | | GLM-4.5-Air-FP8 | 🤗 Hugging Face 🤖 ModelScope | 106B-A12B | FP8 | | GLM-4.5-Base | 🤗 Hugging Face 🤖 ModelScope | 355B-A32B | BF16 | | GLM-4.5-Air-Base | 🤗 Hugging Face 🤖 ModelScope | 106B-A12B | BF16 |

System Requirements#

Inference#

We provide minimum and recommended configurations for "full-featured" model inference. The data in the table below is based on the following conditions:

1. All models use MTP layers and specify --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 to ensure competitive inference speed.
2. The cpu-offload parameter is not used.
3. Inference batch size does not exceed 8.
4. All are executed on devices that natively support FP8 inference, ensuring both weights and cache are in FP8 format.
5. Server memory must exceed 1T to ensure normal model loading and operation.

The models can run under the configurations in the table below:

Under the configurations in the table below, the models can utilize their full 128K context length:

Fine-tuning#

The code can run under the configurations in the table below using Llama Factory:

The code can run under the configurations in the table below using Swift:

Quick Start#

For more comprehensive details and setup instructions, please refer to our GitHub page.

Transformers Inference#

Here is a basic example to run inference with the transformers library, demonstrating both thinking and non-thinking modes:

from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# Load model and tokenizer
model_id = "zai-org/GLM-4.5-FP8"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # Adjust as needed (e.g., torch.float8 for FP8 models)
    low_cpu_mem_usage=True,
    device_map="auto",
    trust_remote_code=True
)
model.eval()

messages = [
    {"role": "user", "content": "Hello, how are you?"},
]

# Example for non-thinking mode (direct response)
# The `add_nothink_token=True` parameter triggers non-thinking mode.
# This mode is suitable for straightforward questions not requiring complex reasoning or tool usage.
inputs_nothink_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, add_nothink_token=True)
input_ids_nothink = tokenizer(inputs_nothink_text, return_tensors="pt").input_ids.to(model.device)
outputs_nothink = model.generate(input_ids_nothink, max_new_tokens=100)
print("Non-thinking mode response:", tokenizer.decode(outputs_nothink[0][len(input_ids_nothink[0]):], skip_special_tokens=True))

# Example for thinking mode (for complex reasoning or tool usage)
# By default, `add_nothink_token=False` or omitting it triggers thinking mode.
# This mode allows the model to perform multi-step reasoning, break down tasks, and utilize tools.
inputs_think_text = tokenizer.apply_chat_template(messages, add_generation_prompt=True, tokenize=False, add_nothink_token=False)
input_ids_think = tokenizer(inputs_think_text, return_tensors="pt").input_ids.to(model.device)
outputs_think = model.generate(input_ids_think, max_new_tokens=100)
print("Thinking mode response:", tokenizer.decode(outputs_think[0][len(input_ids_think[0]):], skip_special_tokens=True))

vLLM#

• Both BF16 and FP8 can be started with the following code:

vllm serve zai-org/GLM-4.5-Air \
    --tensor-parallel-size 8 \
    --tool-call-parser glm45 \
    --reasoning-parser glm45 \
    --enable-auto-tool-choice \
    --served-model-name glm-4.5-air

If you're using 8x H100 GPUs and encounter insufficient memory when running the GLM-4.5 model, you'll need --cpu-offload-gb 16 (only applicable to vLLM).

If you encounter flash infer issues, use VLLM_ATTENTION_BACKEND=XFORMERS as a temporary replacement. You can also specify TORCH_CUDA_ARCH_LIST='9.0+PTX' to use flash infer (different GPUs have different TORCH_CUDA_ARCH_LIST values, please check accordingly).

SGLang#

• BF16

python3 -m sglang.launch_server \
  --model-path zai-org/GLM-4.5-Air \
  --tp-size 8 \
  --tool-call-parser glm45  \
  --reasoning-parser glm45 \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3 \
  --speculative-eagle-topk 1 \
  --speculative-num-draft-tokens 4 \
  --mem-fraction-static 0.7 \
  --served-model-name glm-4.5-air \
  --host 0.0.0.0 \
  --port 8000

• FP8

python3 -m sglang.launch_server \
  --model-path zai-org/GLM-4.5-Air-FP8 \
  --tp-size 4 \
  --tool-call-parser glm45  \
  --reasoning-parser glm45  \
  --speculative-algorithm EAGLE \
  --speculative-num-steps 3  \
  --speculative-eagle-topk 1  \
  --speculative-num-draft-tokens 4 \
  --mem-fraction-static 0.7 \
  --disable-shared-experts-fusion \
  --served-model-name glm-4.5-air-fp8 \
  --host 0.0.0.0 \
  --port 8000

Request Parameter Instructions#

• When using vLLM and SGLang, thinking mode is enabled by default when sending requests. If you want to disable the thinking switch, you need to add the extra_body={"chat_template_kwargs": {"enable_thinking": False}} parameter.
• Both support tool calling. Please use OpenAI-style tool description format for calls.
• For specific code, please refer to api_request.py in the inference folder.

Citation#

If you find our work useful or helpful for your R&D works, please feel free to cite our paper as below.

@article{zhu2025glm45,
  title={GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models},
  author={Zhu, Xiaohan and Sun, Tianxiang and Wang, Hao and Xu, Yi and Zhang, Yichen and Wang, Junyi and Huang, Junjie and Zeng, Jiao and Huang, Yangyang and Gu, Ruipeng and Zhang, Xiaodong and Du, Mengying and Han, Hao and Li, Chao and Xiao, Jin and Guo, Weidong and Li, Zhen and Lu, Jingkang and Chen, Shu and Chen, Huadong and Chen, Peng and Liu, Hongguang and Guo, Guang and Liu, Wen and Yang, Tianyu and Hu, Bo and Zhang, Wenmin and Sun, Maosong},
  journal={arXiv preprint arXiv:2508.06471},
  year={2025}
}