Open-Source GLM-5 Aims for AI Models to 'Become Engineers'
The launch of the open-source model GLM-5 signals a shift toward AI systems that can build and integrate complete software systems, rather than just generating code snippets. This trend suggests a future where agentic large language models could act as design partners and automate parts of the embedded system development toolchain. The model's release reflects a broader movement towards more capable, open-source AI.
- GLM-5 was developed by Zhipu AI, a spinoff from Tsinghua University, and notably, was trained entirely on Huawei Ascend chips without using NVIDIA hardware. The model is a Mixture-of-Experts (MoE) architecture with a total of 744 billion parameters, of which 40-44 billion are active during any given inference operation. - The model's architecture incorporates DeepSeek Sparse Attention (DSA), a technique designed to reduce the computational cost of deployment while maintaining the ability to process long contexts of up to 200,000 tokens. This is a significant increase from its predecessor, GLM-4.5, which scaled to 355 billion parameters. - Zhipu AI is making GLM-5 available as open-source under the permissive MIT License, allowing for unrestricted commercial use and adaptation. The model weights are accessible through platforms like Hugging Face and ModelScope. - In benchmarks for complex, long-duration tasks, GLM-5 has shown strong performance, ranking first among open-source models on Vending Bench 2 by successfully managing a simulated business for a year and achieving a final balance of $4,432. - On coding-specific benchmarks like SWE-bench Verified, GLM-5 achieves a score of 77.8%, and on the multilingual version, it scores 73.3%, outperforming models like GPT-5.2 in the latter. This focus on long-running coding workflows aligns with its positioning for systems-level engineering. - The concept of an "agentic" AI refers to a system that can autonomously set goals, make decisions, and take actions without constant human input, moving beyond simply executing pre-programmed instructions. In embedded systems, this could mean an AI agent managing hardware operations to achieve a goal, like optimizing network connectivity based on real-time conditions. - Integrating AI into embedded systems development introduces challenges related to the non-deterministic nature of large language models, which can complicate verification, traceability, and certification—all critical requirements in safety-conscious fields like the automotive industry. - The push for agentic AI in software engineering is leading to experimental "multi-agentic CI/CD pipelines" where different AI agents collaborate on implementation, compiling, testing, and deployment, automating more of the development lifecycle.
