Z.ai's GLM-5 Model Tops Open-Source Benchmark
What happened
Z.ai has released its GLM-5 model, which has achieved the top score on the open-weights Intelligence Index benchmark. The new model is more than double the size of its predecessor, positioning it as a leading contender in the open-source large language model ecosystem.
Why it matters
- GLM-5 is explicitly designed for "Agentic Engineering," targeting complex, long-horizon tasks that require autonomous planning, tool invocation, and multi-step execution. It has achieved state-of-the-art performance for open-weight models on agentic benchmarks such as BrowseComp (web navigation) and MCP-Atlas (multi-step task execution). - The model's architecture is a Mixture of Experts (MoE) with 744 billion total parameters and 40 billion active parameters per token. It incorporates DeepSeek Sparse Attention (DSA) to efficiently manage its 200K token context window, which is aimed at reducing deployment costs for long-context applications. - Post-training was conducted using a novel asynchronous reinforcement learning (
Key numbers
- Z.ai has released its GLM-5 model, which has achieved the top score on the open-weights Intelligence Index benchmark.
- - GLM-5 is explicitly designed for "Agentic Engineering," targeting complex, long-horizon tasks that require autonomous planning, tool invocation, and multi-step execution.
- The model's architecture is a Mixture of Experts (MoE) with 744 billion total parameters and 40 billion active parameters per token.
- It incorporates DeepSeek Sparse Attention (DSA) to efficiently manage its 200K token context window, which is aimed at reducing deployment costs for long-context applications.
Quick answers
What happened in Z.ai's GLM-5 Model Tops Open-Source Benchmark?
Z.ai has released its GLM-5 model, which has achieved the top score on the open-weights Intelligence Index benchmark. The new model is more than double the size of its predecessor, positioning it as a leading contender in the open-source large language model ecosystem.
Why does Z.ai's GLM-5 Model Tops Open-Source Benchmark matter?
GLM-5 is explicitly designed for "Agentic Engineering," targeting complex, long-horizon tasks that require autonomous planning, tool invocation, and multi-step execution. It has achieved state-of-the-art performance for open-weight models on agentic benchmarks such as BrowseComp (web navigation) and MCP-Atlas (multi-step task execution). The model's architecture is a Mixture of Experts (MoE) with 744 billion total parameters and 40 billion active parameters per token. It incorporates DeepSeek Sparse Attention (DSA) to efficiently manage its 200K token context window, which is aimed at reducing deployment costs for long-context applications. Post-training was conducted using a novel asynchronous reinforcement learning (
