GLM 4.5 Model Release (via)
Super recent model release just from today. Z.ai, yet another Chinese AI lab, dropped another agentic MOE reasoning model. There aren’t any independent benchmarks quite yet, but the benchmarks in their blog post look quite promising. It’s a SOTA level model that’s incredibly cheap and looks to me to be a bit better than Claude 4 Sonnet, which is impressive.
One of the most interesting tidbits that I spotted was that their base pre-training dataset was the exact size (15T tokens) of the FineWeb dataset. This shows to me that the freely available datasets are good enough to get you a SOTA model. For a good base model, the big AI labs don’t have a moat of proprietary data. This is great news for competition. The easier it is to train a SOTA model, the easier it is for new labs to easily get into the game and the more labs we have the more competition there is and the better the models we end up with.
While the General Pre-training Corpus is probably FineWeb, I’m not sure what the Code and Reasoning Pre-training Corpus is. I would guess they used a somewhat similar technique as the DeepSeek R1 training run where DeepSeek used RL to let a regular LLM teach itself how to reason about problems and then finetune a base model on the reasoning traces (Section 2.2 onward in the paper), but I’m not certain if those reasoning traces would be in the 7T pre-training dataset or later in the mid-training.
Anyways, I now know what I’ll be spending my afternoon doing: testing GLM 4.5 performance in Claude Code using a proxy to OpenRouter.