Natural Language Processing

Is GLM-4.5 New Summit Island on Open Source LLM Race?

Automatically translated·See original

Masqot CommunityCreator

August 16, 2025

3 min read

Save

wp:paragrap h Zhipu AI, a Chinese AI initiative, presented the GLM-4.5 AI model, developed for smart artificial intelligence agents! /wp:paragrap h wp:paragrap h So what is this GLM-4.5 initiative? Continue reading to get more details! /wp:paragrap h wp:head g What is GLM-4.5? /wp:headin g wp:paragrap h AI initiative based in China Home Zhipu AI (now Home ) Developed by GLM-4. 5 open source is a great language model with “agent-native” (ajan-oriented) features. It is intended to combine features such as conducting intelligence, control management and integration with external tools in one model. /wp:paragrap h wp:paragrap h GLM-4.5, which has 355 billion parameters, has managed to draw attention from R1 to about 316 million less parameters and GLM-4.5 activates only 32 billion of these parameters at any time to reduce hardware usage. /wp:paragrap h wp:paragrap h Also, thanks to the fact that GLM-4.5 Open-source (Open source) developers can freely use, change or integrate it into other projects in accordance with the MIT license. /wp:paragrap h wp:head g GLM-4.5 Performance and Ranks /wp:headin g wp:paragraph {"className":"} According to the information Z.ai shares; GLM-4.5 in various artificial intelligence criteria Claude 4 Opus Many popular alternatives including left behind. Model in these criteria Grok 4 and OpenAI o3 After that, it was ranked third. /wp:paragrap h wp:paragraph {"className":"} GLM4.5, Home p Tags AIME24 and SWE-Ich Verified Reviews Claude 4 Tags Tags Claude 4 Opus and Ship 2.5 Tags matches or left them behind with leading proprietary models. /wp:paragrap h wp:paragraph {"className":"} Z.ai, also 106 billion parameters GLM-4.5-Air He also developed a reduced version of the model called. About Original Model three times less parameters having GLM-4.5-Air, to process commands Enables 12 billion parameters. /wp:paragrap h wp:head g Technological Innovations /wp:headin g wp:paragrap h GLM-4.5 biggest technical difference Present innovation in two layers: /wp:paragrap h wp:list {"ordered":true} wp:list-item Mixture-of-Experts (MoE) architecture /wp:list-item wp:list-item Hybrid thinking modes (Thinking & Non-thinking) /wp:list-item /wp:list wp:paragrap h These two features reduces calculation cost Home performance optimizes according to task type. /wp:paragrap h wp:heading {"level":3} Mixture-of-Experts (MoE) Architecture /wp:headin g wp:paragrap h MoE actively runs all parameters for each query, a normal giant LLM. GLM-4.5 Although this ensures high accuracy calculating cost ceiling . However, in MoE model, only experts take task and consist of multiple specialist sub models. /wp:paragrap h wp:paragrap h In MoE design in GLM-4.5, the total number of parameters is 355 billion and the number of active parameters in each transaction is about 32 billion. Additionally, up to 80% calculation load reduction offers greater model operation in the same hardware. /wp:paragrap h wp:heading {"level":3} Hybrid Thinking Modes /wp:headin g wp:paragrap h GLM-4.5, according to use two different “ cognitive” mode switch between: /wp:paragrap h wp:list wp:list-item Thinking Mode (Derin Thinking Mode): GLM-4.5 is an effective task model, providing depth in areas such as multi-step reasoning, math problems, etc. /wp:list-item wp:list-item Non-thinking Mode (Fast Response Mode): GLM-4.5 has chatbox scenarios that answer the simple questions from everyday life without losing the conversations that require short response. /wp:list-item /wp:list wp:head g Price Yield Balance /wp:headin g wp:paragrap h One of the most important reasons why GLM-4.5 stands out in the open source LLM market is only performance not, cost optimization u . Designing Z.ai, GLM-4.5 hardware efficiency Home Home API usage costs Home by targeting a “cheap but strong” strategy followed. /wp:paragrap h wp:head g New Summit Island in LLM Race? /wp:headin g wp:paragrap h GLM-4.5 is the top rankings in the ARC tasks in performance evaluation, especially with agentic benchmarks. With Open-Source feature, MIT license and open-weight model provide flexibility and localization advantage. MoE architecture in terms of cost and efficiency of the "Air" variant is more attractive. /wp:paragrap h wp:paragrap h In the light of all these data, GLM-4.5 stands out as a new peak candidate in the open source LLM race. If you want to understand LLM technologies in depth and better evaluate the opportunities offered by GLM-4.5, located on our site LLM (Large Language Model) Technologies You can browse our detailed guide under the header. /wp:paragrap h wp:paragrap h Stay tuned for similar content and developments! /wp:paragrap h wp:paragrap h /wp:paragrap h

Masqot CommunityCreator

@masqot