Neura Market
GLM-130B logo

GLM-130B

Paid

GLM-130B: An open 130B-parameter bilingual transformer for high-performance research and real-world NLP

EducationContact
#bilingual model#transformer-based#NLP#academic research#commercial use#language model#multi-GPU#mixed-precision serving
Inputs: textOutputs: text
Type
Saas
Company
THUDM (Tsinghua University Deep Learning Model Group)
Visit Website

About GLM-130B

GLM-130B is an open-source, bilingual (Chinese and English) general language model developed by THUDM/THUKEG, with 130 billion parameters. It is based on the General Language Model (GLM) algorithm and was trained on over 400 billion tokens, evenly split between Chinese and English. The model is designed for high-performance research and real-world NLP tasks, offering downloadable checkpoints and inference/deployment code for efficient multi-GPU and mixed-precision serving. It supports fast inference on SAT and FasterTransformer, and with INT4 quantization, it can run on a single server with 4 RTX 3090 GPUs with minimal performance loss. The model is released under an open license for academic research and certain commercial use, as detailed in its MODEL_LICENSE.

Key Features

Open bilingual pre-trained model supporting Chinese and English
130B-parameter transformer-based large language model
Open repository with model/code licenses; permitted research and certain commercial uses per repo
Trained on 400B text tokens with the GLM training framework
Optimized large-scale training using the GLM library (parallelism and efficiency)
Strong reported performance across multiple NLP benchmarks
Inference and deployment scripts, including multi-GPU and mixed-precision
Downloadable checkpoints/weights for immediate use
Instruction-style and few-shot prompting capabilities
Associated ICLR 2023 paper describing model design and training

Pros & Cons

Pros
  • Open-source and freely available for academic research and certain commercial use
  • Strong performance on both English and Chinese NLP benchmarks
  • Efficient inference with support for quantization and fast inference frameworks
  • Reduced hardware requirements through INT4 quantization
  • Backed by a well-known research institution (THUDM/THUKEG)
Cons
  • Requires significant computational resources for full-precision inference (e.g., 8 A100 40G GPUs)
  • Free tier limits and commercial use terms should be verified from the MODEL_LICENSE
  • Model size (130B parameters) may be challenging to deploy on consumer hardware without quantization
  • Documentation and community support may be less extensive than some commercial alternatives
  • Performance on specific tasks may vary and should be validated independently

Best For

Academic researchers: Benchmarking bilingual (Chinese–English) NLU and generation tasks with a large, open model.NLP engineers: Building question answering and dialogue systems via prompt-based inference.Data scientists: Rapid prototyping of domain prompts and few-shot workflows without fine-tuning.Enterprise R&D teams: Evaluating 130B-scale model performance for internal pilots under the repository license.Product teams: Drafting bilingual content and templates in Chinese or English for user-facing features.Researchers in efficiency: Studying large-model inference on multi-GPU and mixed-precision setups using provided scripts.Educators: Demonstrating state-of-the-art transformer behavior in classroom or lab settings.Benchmark maintainers: Comparing strong baseline results across standard NLP leaderboards.Applied scientists: Exploring instruction-style prompting and few-shot examples for task completion.Open-source contributors: Extending the GLM library or improving deployment tooling for large models.

Alternatives to GLM-130B

to teach

Transform Your Classroom Using AI-Powered Teaching Resources

LingQ

Create personalized language lessons, engage with interactive activities, and monitor proficiency with quizzes.

Learn Prompting

Unlock the power of AI with Learn Prompting. Learn how to create effective prompts to enhance your AI usage, improve outputs, and solve complex problems.

Reach Best

AI for researching and applying to Universities. Tens of thousands of students use Reach Best to check if a University is a good f

BooksAI

Enhance your reading experience with AI-powered insights

Quizwhiz

Effortlessly Generate Engaging Quizzes and Comprehensive Study Notes with QuizWhiz

FAQ

What hardware is needed to run GLM-130B?
For full-precision inference, GLM-130B can run on a single server with 8×A100 (40G) or 8×V100 (32G). With INT4 quantization, it can run on a single server with 4×RTX 3090 (24G) GPUs with almost no performance loss.
Is GLM-130B free to use?
The model is released under an open license for academic research and certain commercial use. Details are available in the MODEL_LICENSE file in the repository.
What languages does GLM-130B support?
GLM-130B is bilingual, supporting both English and Chinese. It was trained on 200B tokens of each language.
How does GLM-130B compare to other large language models?
GLM-130B outperforms GPT-3 175B (+4.0% on LAMBADA, +0.9% on MMLU), OPT-175B (+5.5%), and BLOOM-176B (+13.0%) on English benchmarks, and significantly outperforms ERNIE TITAN 3.0 260B on Chinese benchmarks.