GLM-5

GLM-5 is an open-weights foundation model aimed at moving programming assistance from casual "vibe coding" toward what its creators call agentic engineering. It is designed to deliver reliable productivity on complex system-building work and long-running agent tasks, with the official documentation positioning it as a state-of-the-art open-source entry on coding and agent benchmarks whose real-world usability approaches that of leading frontier closed models. To support that mission, it adopts DeepSeek Sparse Attention, an architectural optimization that cuts per-token compute and memory pressure while keeping long-context fidelity intact, making it practical to serve a very large model economically rather than just impressive on paper. GLM-5 scales up substantially from its predecessor, jumping to 744 billion parameters with 40 billion active per token, and absorbing a larger 28.5-trillion-token pre-training corpus. The team pairs that scale with a new asynchronous reinforcement learning infrastructure called slime, which decouples generation from training so that fine-grained post-training iterations can run efficiently across long-horizon agent rollouts. The combination of bigger pre-training, sparse attention, and agent-aware RL produces a model that excels at end-to-end software engineering challenges and at autonomous workflows that require sustained planning and self-improvement, making it a strong fit for teams that want open weights and self-hosting control without giving up frontier-style agentic coding capability.

GLM-5 is an open-weights foundation model aimed at moving programming assistance from casual "vibe coding" toward what its creators call agentic engineering. It is designed to deliver reliable productivity on complex system-building work and long-running agent tasks, with the official documentation positioning it as a state-of-the-art open-source entry on coding and agent benchmarks whose real-world usability approaches that of leading frontier closed models. To support that mission, it adopts DeepSeek Sparse Attention, an architectural optimization that cuts per-token compute and memory pressure while keeping long-context fidelity intact, making it practical to serve a very large model economically rather than just impressive on paper.