GLM-5 (Fully Tested): I GOT EARLY ACCESS & YES, IT BEATS 4.6 OPUS!

GLM5: A Detailed Analysis

Key Concepts:

• GLM5: A 744 billion parameter Mixture of Experts (MoE) large language model (LLM).
• Mixture of Experts (MoE): A model architecture where only a subset of the total parameters are activated for each input, increasing efficiency.
• Active Parameters: The number of parameters utilized during a specific inference pass (40 billion in GLM5).
• System Architect Model: An LLM designed for complex system building and agentic engineering, rather than solely focusing on text generation.
• Agentic Engineering: The capability of an LLM to autonomously plan, execute, and debug tasks, acting as an intelligent agent.
• Open Claw: A framework for building AI agents.
• Kilo CLI/Gateway: Tools for interacting with and deploying LLMs.
• Linting: The process of analyzing code for potential errors and stylistic issues.
• Floorplan: A visual representation of a building's layout.
• SVG: Scalable Vector Graphics, an image format.
• 3JS: A JavaScript library for creating 3D graphics.

1. Model Overview & Technical Specifications

GLM5 is a recently released 744 billion parameter LLM utilizing a Mixture of Experts (MoE) architecture. Unlike its predecessor, the GLM4 series (355 billion parameters, 32 billion active), GLM5 activates only 40 billion parameters during usage, contributing to efficiency. While Kimmy, with a trillion parameters, remains the largest open model, GLM5 is now a close competitor. The speaker anticipates a potential increase in API pricing due to the increased parameter count, but expects coding plans to remain at the current price point. Importantly, GLM5 is slated to be released with open weights, and was previously accessible as “pony alpha” on Open Router, with this release being a finalized checkpoint. The model is designed as a “reasoning model” with reasoning tokens available through the API. Initial testing indicates relatively fast performance, though speeds may vary.

2. Shift in Focus: System Architecture vs. Front-End Aesthetics

The developers of GLM5 have explicitly shifted their focus from creating aesthetically pleasing text generators to building a robust “system architect” model. As stated by the developers in 2026, “Coding LLMs are evolving from simply writing code to building systems and GLM5 stands as the first open-source system architect model akin to Claude Opus. Our goal is to shift the narrative from a focus on front-end aesthetics to agentic engineering capabilities.” The speaker strongly agrees with this direction, characterizing GLM5 as comparable to Claude 4.5 or 4.6 Opus, or a combination of Codeex and Opus. This signifies a move towards LLMs capable of complex system design and autonomous task completion.

3. Enhanced Capabilities: Planning, Long-Running Tasks, and Follow-Up Questions

A key weakness of GLM4.7 was its poor planning and debugging capabilities, often skipping steps or failing to grasp the overall architecture of a project. GLM5 addresses this significantly. When used in “plan mode” with tools like Open Code or Kilo Codes, the model effectively checks files, performs system architecture analysis, and proposes detailed plans. Furthermore, GLM5 demonstrates improved ability to ask clarifying questions when prompts are ambiguous, a feature lacking in previous GLM models which would rigidly adhere to initial, potentially flawed, instructions.

The model also excels in long-running tasks, surpassing even recent Opus 4.6 performance in this area. It proactively identifies and fixes linting errors, adheres to user instructions, and consistently works towards the desired outcome.

4. Limitations: Chat and Generative Art

Despite its strengths, GLM5 has limitations. It is not well-suited for general textual chat, performing poorly in this area. It also struggles with generating HTML and SVG images. However, the speaker views these as acceptable trade-offs, given the model’s focus on system architecture and coding. The issue isn’t that GLM5 fails at these tasks, but that it overthinks them due to its extensive training on code and system design, hindering performance on simpler tasks. This behavior is likened to Codeex, which excels at complex tasks but falters on simpler ones.

5. Benchmark Results & Performance Evaluation

The speaker conducted several benchmarks to assess GLM5’s capabilities:

• Floorplan: Scored 10/20 – Functionally sound (toggle roof), but aesthetically lacking.
• Panda with Burger (SVG): Scored 10/20 – Relatively better than previous generations, but not exceptional.
• Pokeball (3JS): Scored 20/20 – Fully functional, interactive (opens and shakes).
• Chessboard with Autoplay: Scored 15/20 – Logical moves, but the board window shifts during piece movement.
• Minecraft Game: Scored 15/20 – Functional, but with minor bugs.
• Butterfly in Garden: Scored highly – Excellent stage setup and overall quality.
• Math Questions: Scored 20/20 – Fully correct answers.

Overall, GLM5 achieved the third position in these benchmarks.

6. Agentic Tests & Real-World Applications

GLM5 demonstrated exceptional performance in agentic tests using Kilo CLI. Notable examples include:

• Movie Tracker App (Expo): Generated a fully functional app in 40 minutes, including linting and front-end error checks using curl commands – significantly better than Opus.
• Go-Based Terminal Calculator: Outperformed both Opus 4.6 and Codeex.
• God Game: Performed well, aligning with current capabilities of other models.
• Spelt to Conban App (Database): Successfully created a functional application.
• Nux App (Stack Overflow Clone): Generated a fully functional clone with threads and posting capabilities, surpassing Opus 4.6.
• Tari App (Image Cropper/Editor with AI): A complex task that took over 3 hours, resulting in a functional, though buggy, application.

7. Agentic Leaderboard & Comparative Analysis

Based on these tests, GLM5 currently holds the number one position on the speaker’s agentic leaderboard. The model consistently completes tasks, produces high-quality code, and avoids premature task completion to save tokens. The speaker now prefers GLM5 over Opus for coding agents, citing its lower cost and overall performance.

8. Accessibility & Cost

GLM5 will be accessible through Open Claw and via the GLM coding plan. The speaker anticipates it will be more affordable than Opus and OpenRouter. He also notes that the developers have made improvements to Kilo CLI to enhance compatibility with GLM5.

9. Notable Quote

“This will probably be the model I use from now on.” – The speaker, expressing his strong preference for GLM5.

Conclusion

GLM5 represents a significant advancement in LLM technology, particularly in the realm of system architecture and agentic engineering. Its enhanced planning capabilities, improved performance in long-running tasks, and focus on practical application make it a compelling alternative to existing models like Claude Opus and Codeex. While it has limitations in areas like chat and generative art, these are considered acceptable trade-offs given its core strengths. The open-weight release and anticipated affordability further solidify GLM5’s potential as a leading LLM for developers and researchers.