GLM-4.5: New SOTA Opensource KING! Powerful, Fast, & Cheap! (Fully Tested)

GLM 4.5 and GLM 4.5 Air: State-of-the-Art Open Source LLMs

Key Concepts: GLM 4.5, GLM 4.5 Air, Large Language Models (LLMs), Mixture of Experts (MoE), Reasoning, Coding, Agentic Capabilities, Context Length, Token Count, Benchmarks, API Pricing, Open Weights, Quantization, SVG Code Generation, AI SaaS Landing Page, Reasoning Prompts, Web Search, AI Slide Generation.

1. Introduction of GLM 4.5 and GLM 4.5 Air

The open-source community has released two new state-of-the-art large language models: GLM 4.5 and GLM 4.5 Air.
These models are designed to unify reasoning, coding, and agentic capabilities.
GLM 4.5: 355 billion total parameters, 32 billion active parameters.
GLM 4.5 Air: 106 billion total parameters, 12 billion active parameters.
Both models support a 128k context length.
Trained on 22 trillion tokens.
Mixture of Experts (MoE) architecture optimized for performance across diverse tasks.

2. Performance and Benchmarking

GLM 4.5 ranked third overall across 12 benchmarks covering agentic tasks, reasoning, and coding.
GLM 4.5 Air ranked sixth.
Competes with models from OpenAI, Anthropic, Google DeepMind, Xi, Alibaba, Moonshot (Kim K2), and DeepSeek.
Features a hybrid thinking mode, switching between deep reasoning/tool use and fast, non-thinking responses.
In reasoning tasks (math, GPQA), GLM 4.5 consistently outperformed or rivaled Claude, GPT-4.1, and Gemini 1.5 Pro.

3. API Pricing and Accessibility

GLM 4.5: $0.60 per 1 million input tokens, $2.20 per 1 million output tokens.
GLM 4.5 Air: $0.20 per 1 million input tokens, $1.10 per 1 million output tokens.
Accessible via the GLM API platform (link in description).
Available through OpenRouter.
Can be used within the Z AI chatbot.
Open weights are available (link in description).
Quantized versions will be released through Ollama or LM Studio for local use.

4. Coding Capabilities and Examples

Demonstrated ability to create a functional Flappy Bird game in one shot.
Generates UI and functionality accurately.
Examples include a to-do board, a timeline, PowerPoint presentations, and basic frontends.
Performs well in front-end development, demonstrated by a Pokémon catalog with animations.
Considered better than the Kim K2 model.

5. Testing and Benchmarking Examples

SVG Code Generation: Asked to create a butterfly in SVG code. The model successfully generated a complex and aesthetically pleasing butterfly, showcasing its code generation ability, geometry, and spatial reasoning. This was previously a difficult task for open-source models.
AI SaaS Landing Page: Used Z AI's full-stack feature to create an AI resume app landing page. The model generated a visually appealing front-end with basic animations and styles, utilizing technologies like Next.js and JS Chadients.
Hard Reasoning Prompt: Presented a detective scenario to evaluate logical deduction. The model correctly identified the thief (Bob) and the liar (Dana), providing a detailed explanation of its reasoning process.

6. Z AI Chatbot Tools

The GLM chatbot includes tools like web search, slides maker, workspace, and image search.
Web Search: Example: "What was the most recent Nvidia closing price?" The model accurately retrieved the price ($176.75 as of 4:00 PM).
AI Slide Generation: Example: "Create a slide on the World of AI YouTube channel." The model generated a slide deck with accurate subscriber count, video count, channel description, content categories, popular videos, trending topics, community engagement, and related channels.

7. Conclusion

GLM 4.5 and GLM 4.5 Air are impressive open-source models excelling in reasoning, coding, and agentic tasks.
They rival or surpass other models like Claude 3 Opus in various categories.
The models are highly recommended for experimentation and use.
The presenter encourages viewers to explore the models using the provided links, subscribe to the channel, join the newsletter and Discord, and follow on Twitter.