🚀 Just dropped: the GLM-4.5 tech report! This powerhouse Mixture-of-Experts (MoE) LLM packs 355 billion parameters (32 billion active) and has hybrid logic—think "deep thoughts" for tricky tasks and quick responses when it counts! 🤯
Key features:
- Trained on a whopping 23 trillion tokens with search fine-tuning & RL through expert iterations.
- Excels at agentic tasks, reasoning, and coding:
— TAU-Bench: 70.1% ✔️
— AIME 24: 91.0% 🎉
— SWE-bench Verified: 64.2% 💻
- Despite its size, it's ranked #3 in overall metrics and #2 in agentic benchmarks among all models! 🥉🥈
- Two versions out: full-size GLM-4.5 (355B) and compact GLM-4.5-Air (106B)—both open to the community! 🌍
This is a game-changer for open LLMs—a hybrid champ that can reason, act, and code—all in one framework! 💪✨
