GLM-4.5 (Fully Tested) + Free APIs + Cline,Roo,Kilo: This is OPEN GPT-5! Better than Kimi-K2?

Key Concepts:

• GLM 4.5 & GLM 4.5 Air: Large language models (LLMs) by ZAI/JIPU AI.
• Mixture of Experts (MoE): Architecture used to build GLM 4.5 and GLM 4.5 Air, allowing for both power and efficiency.
• Agent Native Design: Models trained with built-in agent capabilities like tool calling.
• Hybrid Reasoning: Ability of the model to toggle thinking processes.
• Kilo Code, Klein, RU Code, Claude Code: IDE extensions for utilizing LLMs in coding.
• Open Router: Platform for accessing various LLMs, including GLM 4.5.
• ZAI API: Direct API access to ZAI's models.
• Token Pricing: Cost per million input and output tokens for using the models.

1. Introduction to GLM 4.5 and GLM 4.5 Air

• ZAI/JIPU AI has launched GLM 4.5 and GLM 4.5 Air. This company was previously identified by OpenAI as a potential major player in the AI market from China.
• GLM 4.5 is the flagship model with 355 billion parameters.
• GLM 4.5 Air is a lighter version with 106 billion parameters, designed for lower hardware demands and faster inference.
• Both models are built on a fully self-developed mixture of experts architecture which is kind of cool because it allows them to be both powerful and efficient.
• ZAI positions GLM 4.5 as China's most advanced open-source model, fully open, auditable, and suitable for on-premise deployment.

2. Performance and Capabilities

• GLM 4.5 is presented as a competitive open-source model, topping domestic Chinese model charts.
• GLM 4.5 Air leads in performance and efficiency among models in the 100B parameter range.
• Both models have "agent native design," meaning they are trained with agent capabilities such as tool calling, enabling integration with platforms like Kilo Code, Klein, and RU Code.
• Hybrid reasoning capabilities are available, where the model can toggle between thinking or not thinking.
• Based on the average scores across these benchmarks, GLM 4.5 actually ranks pretty well, which is honestly impressive if it's true.

3. Pricing and Speed

• GLM 4.5 is competitively priced via the official Chinese API at 11 cents per million input tokens and 28 cents per million output tokens.
• It can generate over 100 tokens per second, providing a balance between speed and cost.
• Open Router offers access to GLM 4.5 via Singapore servers at a higher cost: 60 cents per million input and $2.20 per million output tokens, and $1.10 for the Air model.

4. Practical Usage and Integration

• VS Code Integration: Users can integrate GLM 4.5 and GLM 4.5 Air using extensions like Kilo Code, Klein, and RU Code.
  • Kilo Code offers $20 of free credits for trying out the models.
  • Klein allows selecting Open Router and the GLM model.
  • Users can also use the ZAI API directly by providing the API key and base URL.
• Claude Code Integration: GLM 4.5 can be integrated with Claude Code by exporting the ZAI API key, base URL, and O token.

5. User Experience and Benchmarking

• The presenter tested GLM 4.5 on benchmark questions, achieving a score of two out of five with Claude Code, similar to Quen Code.
• The presenter noted that GLM 4.5 is similar to Quen 3 Coder and Kim K2, but is smaller and cheaper.
• The model can be verbose in its outputs, and the hybrid thinking feature might not appeal to all users.
• The Air model is suitable for local use on high-tier MacBooks due to its ANE (Apple Neural Engine) optimization.

6. Sponsor: Dart

• Dart combines traditional project management with AI features.
• Dart's AI can brainstorm project ideas, generate task lists, and complete assignments.
• Custom agents can be created to trigger from built-in integrations or custom web hooks, enabling automated tasks like coding pull requests, marketing campaigns, and email outreach.
• Dart integrates with Claude, Chat GPT, and other AI tools through their MCP server.
• Most features are free, with premium options starting at $8 per month.

7. Conclusion

GLM 4.5 and GLM 4.5 Air from ZAI/JIPU AI present themselves as powerful, open-source language models with competitive pricing and agent capabilities. The models offer flexibility in deployment and integration, with options for cloud-based access through Open Router or direct API usage. While the models have some limitations, such as verbosity in outputs, their performance and cost-effectiveness make them attractive options for various applications, especially for users looking for alternatives to established models like Quen and Kimmy.

Key Takeaways:

• GLM 4.5 is a large, open-source model with 355B parameters, positioning itself as a competitor to other leading models.
• GLM 4.5 Air offers a lighter alternative with 106B parameters, suitable for local use and faster inference.
• The models feature agent-native design and hybrid reasoning capabilities.
• Pricing is highly competitive, particularly through the official Chinese API.
• Integration with popular coding environments like VS Code is streamlined through extensions like Kilo Code, Klein, and RU Code.