GLM-5 Leaked? New Pony Alpha Stealth Model IS INSANE! Opus 4.6 Quality!

Pony Alpha & GLM5: A Deep Dive into a Cutting-Edge Open-Source Model

Key Concepts:

• Pony Alpha: A new stealth model on Open Router, strongly suspected to be GLM5.
• GLM5: A rumored large language model (LLM) from Zai Corp, potentially the largest Chinese Mixture of Experts (MoE) model.
• Mixture of Experts (MoE): A neural network architecture where different parts of the model specialize in different tasks.
• Context Window: The amount of text a model can consider at once when generating output (Pony Alpha has a 200K context window).
• Open Router/Arena/Hilo: Platforms for testing and accessing various LLMs.
• Kilo: An API provider offering access to Pony Alpha and an AI open-source coding agent.
• SVG: Scalable Vector Graphics, a format for defining vector-based images.
• 3.js: A JavaScript library for creating and displaying animated 3D computer graphics in a web browser.
• DeepSeek v3: A competing LLM used for performance comparison.
• Opus 4.5/4.6: Models from Anthropic, used as benchmarks for Pony Alpha’s output quality.

---

I. Introduction of Pony Alpha & GLM5

A new, currently free-to-access stealth model, “Pony Alpha,” is available on Open Router and is widely believed to be the upcoming GLM5 model developed by Zai Corp. Initial testing, including by the video creator, indicates strong performance in coding, agentic workflows, reasoning, and text generation, making it suitable for real-world coding applications. Pony Alpha currently ranks 10th on Open Router’s programming leaderboard. Checkpoints are being tested across platforms like Open Router, Arena, and Hilo, with the Transformer code for GLM5 spotted on Zi’s GitHub, suggesting a release window within February.

II. GLM5 Specifications & Rumors

Rumors suggest GLM5 will boast 745 billion parameters, with 44 billion active parameters, exceeding the size of DeepSeek version 3. This would make it the largest Chinese Mixture of Experts (MoE) model to date. Leaks from Hugging Face Transformers indicate GLM5’s total parameter size is approximately twice that of GLM 4.5, incorporating DC sparse attention for efficient handling of long context. This architecture is expected to significantly enhance its long-context capabilities and generation quality. The creator notes the output quality is, in their opinion, comparable to Opus 4.5, though this is acknowledged as a debated perspective.

III. Model Identification & Provider Confirmation

Interestingly, when the system prompt on Open Router is changed to “custom” and left blank, Pony Alpha identifies itself as being provided by Zai and confirms it is a GLM model. This further supports the claim that Pony Alpha is indeed GLM5.

IV. Concerns Regarding Training Data & Origins

The video touches upon a common discussion within the AI community: the potential reliance of Chinese AI companies on the outputs of American firms like Anthropic and Google for training their models. The creator states that companies like Miniax and ZAI Corp likely utilize this approach, though details are often not publicly shared. Access to more in-depth information is available through the creator’s free newsletter (link in description).

V. Accessing & Utilizing Pony Alpha & Aurora Alpha

Both Pony Alpha and another recently released stealth model, “Aurora Alpha” (potentially an OpenAI model mimicking GPT-OSS), are accessible via Open Router and Arena. For API access, Kilo provides free access to Pony Alpha, along with a powerful AI open-source coding agent. The creator demonstrates using Kilo within VS Code, highlighting its ease of installation and use.

VI. Practical Demonstrations & Performance Analysis

The video showcases several practical demonstrations of Pony Alpha’s capabilities using Kilo:

• SVG Generation: The model successfully generated SVG code for a butterfly, with the second iteration including animation and a photorealistic appearance. While not surpassing Opus 4.6, the SVG generation quality is considered impressive. Examples of other SVG scenes generated by Pony Alpha are also shown.
• Frontend Development: Pony Alpha generated a complete, interactive landing page, demonstrating its ability to structure layouts, wire up dynamic elements, and incorporate smooth animations. The creator notes that a non-programmer would likely be unable to distinguish it from a human-created page.
• AI Agent-Powered Frontend: Using Kilo’s AI agent, the model generated a structured landing page for a celebrity, showcasing dynamic behavior and styling.
• Browser-Based OS: Pony Alpha created a functional, browser-based OS ("Pony OS") with various applications, including a web browser, settings panel (with a non-functional dark mode), weather app, and even a Minesweeper game. The OS design draws inspiration from both macOS and Windows.
• Minecraft Clone: The model attempted to generate a Minecraft clone using 3.js. The initial version had issues, but a second iteration showed improvements in terrain gradients and functionality, allowing for block breaking and placement.

VII. Performance & Speed Comparison

The creator observes that Pony Alpha generates code faster than many other models while maintaining a high level of quality in its generations.

VIII. Conclusion & Future Outlook

The GLM5 (Pony Alpha) is anticipated to be a significant open-source model. The creator expects a release within February, with further refinement of the output quality likely. They draw a parallel to the Miniax model, where early testing was followed by output optimization. The video encourages viewers to explore the model through the provided links, subscribe to the creator’s channels and newsletter, and stay updated on the latest AI news.

Notable Quote:

“It’s comparable to the output of Opus 4.5. And this is my opinion, so don't come at me. I just personally believe that the output is similar to the Opus 4.5.” – The video creator, regarding Pony Alpha’s output quality.