Unveiling GLM-5.1: A Leap in Agentic Engineering and Coding Capabilities

GLM-5.1 represents a significant advancement in agentic engineering, boasting superior coding capabilities over its predecessor, GLM-5. It excels in state-of-the-art performance across various benchmarks, including SWE-Bench Pro, NL2Repo, and Terminal-Bench 2.0. Unlike previous models that plateau after initial gains, GLM-5.1 maintains effectiveness over extended periods, handling complex and ambiguous tasks with precision.

## Complex Software Engineering Tasks

GLM-5.1 shines in complex software engineering tasks, surpassing other models like GPT-5.4 and Gemini 3.1 Pro on SWE-Bench Pro. It doesn't just perform well initially but continues to optimize over time, breaking down problems, running experiments, and iterating strategies to sustain improvements over hundreds of iterations.

### Scenario 1: Vector Database Optimization

In a test involving the VectorDBBench challenge, GLM-5.1 demonstrated its prowess by optimizing a vector database over 600 iterations, achieving a QPS of 21.5k—six times the best result in a single session. This was achieved through strategic transitions and self-analysis, showcasing its ability to identify and overcome bottlenecks.

### Scenario 2: Machine Learning Workload Optimization

GLM-5.1 also excels in optimizing machine learning workloads, achieving a 3.6× speedup in GPU kernel tasks. While its rate of improvement slows over time, it sustains optimization longer than GLM-5, although Claude Opus 4.6 remains the top performer in this domain.

### Scenario 3: Building a Linux Desktop

In a more subjective task of building a Linux desktop environment as a web application, GLM-5.1 continued to refine and enhance the application over 8 hours, resulting in a complete and polished desktop environment. This demonstrates its ability to self-evaluate and improve without explicit metrics.

GLM-5.1's extended productive horizon highlights the importance of runtime in achieving optimal results. It opens new possibilities in long-horizon optimization, although challenges remain in escaping local optima and maintaining coherence over extensive execution traces. Released under the MIT License, GLM-5.1 is available on various platforms and supports local deployment, offering developers a powerful tool for complex coding tasks.