Is GLM‑4‑9B the New Powerhouse? A Deep Dive into Its Performance and Usage

Model Overview

GLM-4-9B is an open‑source 9‑billion‑parameter large language model released by Zhipu AI. It is part of the GLM‑4 series and provides three main families:

GLM‑4‑9B – base model, 8K token context.

GLM‑4‑9B‑Chat – chat‑tuned, 128K token context.

GLM‑4‑9B‑Chat‑1M – chat‑tuned, 1 million token context.

GLM‑4V‑9B – multimodal variant (vision + language), 8K token context.

The models are released under an open‑source license, with code and weights hosted at:

https://github.com/THUDM/GLM-4/tree/main

They support Retrieval‑Augmented Generation (RAG), tool‑calling, and up to 1 M tokens of context, enabling “needle‑in‑a‑haystack” retrieval experiments with 100 % success.

Installation and Basic API Call

Install the official SDK and query the model via the REST API:

# pip install zhipuai
from zhipuai import ZhipuAI
client = ZhipuAI(api_key="YOUR_API_KEY")
response = client.chat.completions.create(
    model="glm-4-9b",
    messages=[
        {"role": "system", "content": "You are a helpful assistant."},
        {"role": "user", "content": "你好"}
    ],
    top_p=0.7,
    temperature=0.95,
    max_tokens=1024,
    tools=[{"type": "web_search", "web_search": {"search_result": true}}],
    stream=True,
)
for chunk in response:
    print(chunk)

Quick‑Start Inference – Transformers Backend

Run the model locally with transformers and torch (single‑GPU recommended). The example uses the chat‑tuned checkpoint THUDM/glm-4-9b-chat:

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda"

tokenizer = AutoTokenizer.from_pretrained(
    "THUDM/glm-4-9b-chat", trust_remote_code=True
)
query = "你好"
inputs = tokenizer.apply_chat_template(
    [{"role": "user", "content": query}],
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt",
    return_dict=True,
)
inputs = inputs.to(device)
model = AutoModelForCausalLM.from_pretrained(
    "THUDM/glm-4-9b-chat",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
).to(device).eval()

gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
with torch.no_grad():
    outputs = model.generate(**inputs, **gen_kwargs)
    outputs = outputs[:, inputs["input_ids"].shape[1]:]
    print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Quick‑Start Inference – vLLM Backend

For higher throughput, use the vllm engine. The following script runs the 128K‑context chat model:

from transformers import AutoTokenizer
from vllm import LLM, SamplingParams

max_model_len, tp_size = 131072, 1  # 128K tokens, single GPU
model_name = "THUDM/glm-4-9b-chat"
prompt = "你好"

tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
llm = LLM(
    model=model_name,
    tensor_parallel_size=tp_size,
    max_model_len=max_model_len,
    trust_remote_code=True,
    enforce_eager=True,
)
stop_token_ids = [151329, 151336, 151338]
sampling_params = SamplingParams(temperature=0.95, max_tokens=1024, stop_token_ids=stop_token_ids)
inputs = tokenizer.build_chat_input(prompt, history=None, role='user')["input_ids"].tolist()
outputs = llm.generate(prompt_token_ids=inputs, sampling_params=sampling_params)
print([output.outputs[0].text for output in outputs])

Multimodal Inference (GLM‑4V‑9B)

GLM‑4V‑9B adds vision capabilities. Example using transformers:

import torch
from PIL import Image
from transformers import AutoModelForCausalLM, AutoTokenizer

device = "cuda"

tokenizer = AutoTokenizer.from_pretrained("THUDM/glm-4v-9b", trust_remote_code=True)
query = "描述这张图片"
image = Image.open("your_image.jpg").convert('RGB')
inputs = tokenizer.apply_chat_template(
    [{"role": "user", "image": image, "content": query}],
    add_generation_prompt=True,
    tokenize=True,
    return_tensors="pt",
    return_dict=True,
)
inputs = inputs.to(device)
model = AutoModelForCausalLM.from_pretrained(
    "THUDM/glm-4v-9b",
    torch_dtype=torch.bfloat16,
    low_cpu_mem_usage=True,
    trust_remote_code=True,
).to(device).eval()

gen_kwargs = {"max_length": 2500, "do_sample": True, "top_k": 1}
with torch.no_grad():
    outputs = model.generate(**inputs, **gen_kwargs)
    outputs = outputs[:, inputs["input_ids"].shape[1]:]
    print(tokenizer.decode(outputs[0]))

Benchmark Highlights

GLM‑4‑9B‑Chat consistently outperforms Llama‑3‑8B‑Instruct on a wide range of dialogue and reasoning benchmarks, approaching GPT‑4‑Turbo performance.

AlignBench : 7.01 (GLM‑4‑9B‑Chat) vs. 6.40 (Llama‑3‑8B‑Instruct)

MT‑Bench : 8.35 vs. 8.00

HumanEval : 71.8 vs. 62.2

MMLU (base) : 74.7 vs. 66.6

C‑Eval (base) : 77.1 vs. 51.2

GSM8K (base) : 84.0 vs. 45.8

Multilingual evaluation on six datasets shows GLM‑4‑9B‑Chat leading Llama‑3‑8B‑Instruct:

M‑MMLU: 56.6 vs. 49.6

FLORES: 28.8 vs. 25.0

XStoryCloze: 90.7 vs. 84.7

Tool‑calling performance on the Berkeley Function Calling Leaderboard:

Overall accuracy: 81.00 (GLM‑4‑9B‑Chat) – comparable to GPT‑4‑Turbo (81.24) and far above ChatGLM3‑6B (57.88).

Multimodal visual‑language benchmarks (MMBench, SEEDBench_IMG, MMStar, MMMU, MME, HallusionBench, AI2D, OCRBench) place GLM‑4V‑9B on par with or above leading models such as GPT‑4o and InternVL‑Chat‑V1.5.

Evaluation Script Example

For the 1 M‑token “needle‑in‑a‑haystack” experiment, the evaluation script is available at:

https://github.com/LargeWorldModel/LWM/blob/main/scripts/eval_needle.py

GLM‑4‑9B was pretrained with a mixture of mathematics, reasoning, and code instruction data, which is why it is compared against Llama‑3‑8B‑Instruct.

Key Takeaways

Despite its modest 9 B parameter count, GLM‑4‑9B delivers competitive performance across dialogue, multilingual, tool‑calling, and multimodal tasks, while supporting single‑GPU deployment and extremely long context windows (up to 1 M tokens). This makes it a practical choice for researchers and developers seeking high‑quality open‑source LLMs without large hardware requirements.