Paper summary
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.
- The model targets three things at once, agentic, reasoning and coding tasks, which the authors call "ARC".
- It ranks 3rd overall and 2nd on agentic benchmarks among the models evaluated, with fewer parameters than several competitors.