Login
Login

Draft:GLM

{{AFC submission|d|v|u=Mrfakename|ns=118|decliner=BuySomeApples|declinets=20250429074515|reason2=web|ts=20250429004958}}

{{Short description|Family of large language models by Zhipu AI and THUDM}}

{{Draft topics|internet-culture|software|computing|technology}}

{{AfC topic|other}}

{{Infobox software

| title = GLM

| logo =

| logo_upright =

| screenshot =

| caption =

| developer = Zhipu AI

| released = {{Start date and age|2022|8}}

| latest release version = 4-0414

| latest release date = {{start date and age|2025|4}}

| repo = {{URL|https://github.com/THUDM/GLM-4}}

| programming language = Python

| operating system = {{Plainlist|

}}

| genre = Chatbot

| license = MIT, ChatGLM License, GLM-4 License

| website = {{URL|https://chat.z.ai/}}

}}

GLM (also called ChatGLM or General Language Model) is a family of open-source bilingual Large Language Models developed by Z.ai (Zhipu AI) and The Knowledge Engineering Group & Data Mining (THUDM) at Tsinghua University. In April 2025, the latest GLM model family, GLM-4-0414, achieved competitive performance on benchmarks with leading models such as OpenAI's GPT-4o and DeepSeek's R1.{{Cite web |last=Razzaq |first=Asif |date=2025-04-14 |title=THUDM Releases GLM 4: A 32B Parameter Model Competing Head-to-Head with GPT-4o and DeepSeek-V3 |url=https://www.marktechpost.com/2025/04/14/thudm-releases-glm-4-a-32b-parameter-model-competing-head-to-head-with-gpt-4o-and-deepseek-v3/ |access-date=2025-04-28 |website=MarkTechPost |language=en-US}}

History

The first version of GLM, GLM-130B, was released in August 2022.{{Cite web |title=GitHub - THUDM/GLM-130B at 99509f09cce5ac32aee8ebf78e18f4399319dba4 |url=https://github.com/THUDM/GLM-130B/tree/99509f09cce5ac32aee8ebf78e18f4399319dba4 |access-date=2025-04-29 |website=GitHub |language=en}} While not open-sourced, the model was made available to researchers through an application for non-commercial research purposes.

The first publicly-available open-access GLM model was released in February 2023 with 2B and 10B variants, designed for both blank infilling and text generation.{{Cite web |title=THUDM/glm-10b |url=https://huggingface.co/THUDM/glm-10b |access-date=2025-04-28 |website=huggingface.co}}

= ChatGLM =

In March 2023, ChatGLM 6B was open-sourced on Hugging Face as a conversational language model inspired by ChatGPT.{{Cite web |title=THUDM/chatglm-6b |url=https://huggingface.co/THUDM/chatglm-6b |access-date=2025-04-28 |website=huggingface.co}} In May 2023, VisualGLM, a 6B parameter vision language model based on ChatGLM, was released.

In June 2023, ChatGLM 2 6B was released with improved performance and a variant with a 32K context window.{{Cite web |title=THUDM/chatglm2-6b |url=https://huggingface.co/THUDM/chatglm2-6b |access-date=2025-04-29 |website=huggingface.co}}

In October 2023, ChatGLM 3 6B was released with both chat and base variants, with added support for function calling and improved performance.{{Cite web |title=THUDM/chatglm3-6b |url=https://huggingface.co/THUDM/chatglm3-6b |access-date=2025-04-29 |website=huggingface.co}}

= GLM-4 =

In June 2024, the GLM-4 model family was released with 9 billion parameters. The model family includes GLM-4-Chat, GLM-4-Chat-1M (with a 1M token context window), and GLM-4V-9B (a vision language model). The model was competitive with other open models, outperforming Mistral 7B but falling short of contemporaneous commercial models such as GPT-4 and Claude 3 Opus.{{Cite web |date=2024-08-12 |title=THUDM/glm-4-9b-chat |url=https://huggingface.co/THUDM/glm-4-9b-chat/blob/main/README_en.md |access-date=2025-04-29 |website=huggingface.co}}

= GLM-4-0414 =

In April 2025, Zhipu AI released GLM-4-0414, a model based on GLM-4 with greatly improved performance. The model was trained on 15T tokens of web and synthetic data, with performance competitive with DeepSeek V3 and GPT-4o.{{Cite web |date=2025-04-15 |title=THUDM/GLM-4-32B-0414 |url=https://huggingface.co/THUDM/GLM-4-32B-0414 |access-date=2025-04-29 |website=huggingface.co}}{{Cite web |last=Razzaq |first=Asif |date=2025-04-14 |title=THUDM Releases GLM 4: A 32B Parameter Model Competing Head-to-Head with GPT-4o and DeepSeek-V3 |url=https://www.marktechpost.com/2025/04/14/thudm-releases-glm-4-a-32b-parameter-model-competing-head-to-head-with-gpt-4o-and-deepseek-v3/ |access-date=2025-04-29 |website=MarkTechPost |language=en-US}} The following variants were made available:

  • GLM-4-32B-0414: A chat model with 32B parameters fine-tuned from the GLM-4 Base 0414 model.{{Cite web |date=2025-04-15 |title=THUDM/GLM-4-32B-0414 |url=https://huggingface.co/THUDM/GLM-4-32B-0414 |access-date=2025-04-29 |website=huggingface.co}}
  • GLM-4-32B-Base-0414: A base model with 32B parameters.{{Cite web |date=2025-04-15 |title=THUDM/GLM-4-32B-Base-0414 |url=https://huggingface.co/THUDM/GLM-4-32B-Base-0414 |access-date=2025-04-29 |website=huggingface.co}}
  • GLM-Z1-32B-0414: A reasoning model similar to OpenAI's o1 based on GLM-4.{{Cite web |date=2025-04-15 |title=THUDM/GLM-Z1-32B-0414 |url=https://huggingface.co/THUDM/GLM-Z1-32B-0414 |access-date=2025-04-29 |website=huggingface.co}}
  • GLM-Z1-32B-Rumination-0414: A reasoning model based on GLM-Z1-32B but with support for "rumination," similar to OpenAI's Deep Research.{{Cite web |date=2025-04-15 |title=THUDM/GLM-Z1-Rumination-32B-0414 |url=https://huggingface.co/THUDM/GLM-Z1-Rumination-32B-0414 |access-date=2025-04-29 |website=huggingface.co}}
  • GLM-4-9B-0414: A chat model with 9B parameters.{{Cite web |date=2025-04-15 |title=THUDM/GLM-4-9B-0414 |url=https://huggingface.co/THUDM/GLM-4-9B-0414 |access-date=2025-04-29 |website=huggingface.co}}
  • GLM-Z1-9B-0414: A variant of the GLM-Z1 model based on the 9B chat model.{{Cite web |date=2025-04-15 |title=THUDM/GLM-Z1-9B-0414 |url=https://huggingface.co/THUDM/GLM-Z1-9B-0414 |access-date=2025-04-29 |website=huggingface.co}}

Notably, the GLM-4-9B-0414 base model was not released.{{Cite web |date=2025-04-14 |title=THUDM/GLM-4-9B-0414 · Base model |url=https://huggingface.co/THUDM/GLM-4-9B-0414/discussions/1 |access-date=2025-04-29 |website=huggingface.co}}

Licenses

The GLM and ChatGLM models have been licensed under a wide variety of licenses. The original GLM-130B model is licensed under a non-commercial license prohibiting commercial use.{{Cite web |title=GLM-130B Application Form |url=https://docs.google.com/forms/d/e/1FAIpQLSehr5Dh_i3TwACmFFi8QEgIVNYGmSPwV0GueIcsUev0NEfUug/viewform |access-date=2025-04-29 |website=Google Docs |language=en}} The GLM 2B and 10B did not include a license for the model, but the code itself was licensed under the MIT license.{{Citation |title=THUDM/GLM |date=2025-04-27 |url=https://github.com/THUDM/GLM |access-date=2025-04-29 |publisher=Z.ai & THUKEG}} The ChatGLM models are licensed under a restrictive revocable license.{{Cite web |date=2023-07-08 |title=MODEL_LICENSE · THUDM/chatglm-6b at main |url=https://huggingface.co/THUDM/chatglm-6b/blob/main/MODEL_LICENSE |access-date=2025-04-29 |website=huggingface.co}} The GLM-4 models are licensed under a revocable license that requires attribution.{{Cite web |date=2024-06-08 |title=LICENSE · THUDM/glm-4-9b at main |url=https://huggingface.co/THUDM/glm-4-9b/blob/main/LICENSE |access-date=2025-04-29 |website=huggingface.co}} The GLM-4-0414 model family is licensed under the MIT license, a permissive, OSI-approved license.{{Cite web |date=2025-04-15 |title=THUDM/GLM-4-32B-0414 |url=https://huggingface.co/THUDM/GLM-4-32B-0414 |access-date=2025-04-29 |website=huggingface.co}}

References

{{reflist}}