Paper summary: GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models

GLM-4.5 is an open MoE LLM (355B total, 32B active) with hybrid thinking and direct modes, trained on 23T tokens plus expert iteration and RL, reporting 70.1% TAU-Bench, 91.0% AIME 24 and 64.2% SWE-bench Verified.

Figure from the paper "GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models"

Figure from Team et al. (2025), licensed CC BY 4.0. Figure 1 : Average performance on agentic, reasoning, and coding (ARC) benchmarks. Overall, GLM-4.5 achieves a rank of 3rd, with GLM-4.5-Air following at rank 6th. The models listed are evaluated as of July 28, 2025.

Problem

Strong performance on agent tasks, reasoning and coding in one open model, at a manageable active size.

Method

GLM-4.5 is a Mixture-of-Experts model with 355B total and 32B activated parameters. It supports both thinking and direct-response modes. Training uses 23T tokens in multiple stages, then post-training with expert model iteration and reinforcement learning. A compact GLM-4.5-Air (106B) is released too.

Key results

From the abstract:

70.1% on TAU-Bench (agentic)
91.0% on AIME 24 (maths)
64.2% on SWE-bench Verified (coding)
ranks 3rd overall among all evaluated models and 2nd on agentic benchmarks (Figure 1)

Limitations

There’s no dedicated limitations section in the extracted text. The ranking depends on which models and benchmarks the authors chose to compare.

Paper summary: GLM-4.5: Agentic, Reasoning, and Coding (ARC) Foundation Models

Problem

Method

Key results

Limitations

Sources