The video presents a first look at GLM-5.1, an open-source coding model recently released by Z.A.I. This model is notable for its high performance on various software engineering benchmarks, positioning it slightly below top-tier frontier models like GPT-4 but ahead of popular open-source coding models such as Quen 3.6 and Miniax M27. Unlike the previously reviewed GLM 5V Turbo, which focused on multimodal visual coding, GLM-5.1 is designed specifically for coding tasks and autonomous agentic coding, excelling in long-horizon projects like building a Linux desktop from scratch. It is available on platforms like Hugging Face and Open Router, with competitive API pricing and the option to run locally for free, though its large size (7.54 billion parameters) makes local use challenging for many users.
The presenter demonstrates GLM-5.1’s capabilities by integrating it with a custom codebase focused on adapting sparse attention methods. Using the Hermes agent, the model reviews the codebase, researches relevant papers, and suggests five specific improvements to enhance speed and performance. The suggestions are detailed and reference exact lines of code, showing the model’s ability to conduct meaningful research and provide actionable feedback. While there was one hallucinated suggestion, the overall quality of the review was strong, indicating the model’s potential for assisting in complex coding tasks and iterative development.
Next, the presenter tasks GLM-5.1 with a more complex, long-horizon project: analyzing lengthy trading journal transcripts to extract an executable trading strategy and then build a live screening monitor and paper trading system using APIs like Hyperliquid. Despite some initial tooling hiccups and slower processing compared to models like Opus 4.6, GLM-5.1 successfully generates Python scripts for data feeds, configuration, and foundational components of the trading system. It plans out additional modules such as the trading signal engine and live screener, demonstrating its ability to handle multi-step coding projects with minimal prompting.
After about an hour of processing, GLM-5.1 completes the trading system, including a 3-day backtest that shows a 37% win rate but a strong profit due to high reward-to-risk trades, consistent with the ICT trading strategy principles. The model’s autonomous approach requires less user intervention than other models, making it effective for complex tasks despite slower speed. The presenter highlights the model’s ability to distill complex information from transcripts into functional code and its practical application in real-time trading scenarios, showcasing its strength in both research and coding execution.
In conclusion, GLM-5.1 stands out as a powerful open-source coding model that balances performance, cost, and accessibility. While it may not match the speed of some proprietary models, its quality and autonomous capabilities make it a compelling choice for developers and researchers. The open-source nature encourages community-driven optimization and adaptation for smaller hardware. The presenter invites viewers to share their experiences with GLM-5.1 and expresses intent to continue reviewing significant new models as they emerge, emphasizing the rapid pace of AI development in coding applications.
