GLM-5.2 Scores 99.2% on AIME With 40B Params

GLM-5.2 hit 99.2% on the AIME 2026 math benchmark using 40 billion active parameters, a fraction of GPT-4's roughly 280 billion in 2023, per Walter van der Giessen.

The trade-off matters for builders choosing a model: a 40B-parameter reasoner now beats systems seven times its size on math.

Qwen3.5 scores 91.3% on the same test with 17 billion active parameters.

Giessen's explanation is capacity math: storing facts costs about two bits of knowledge per parameter, and facts go stale. Reasoning skills like decomposition, verification and backtracking compress into far fewer parameters and don't expire.

The catch: the best-performing model on SimpleQA hits only 53% factual recall. Qwen's 4B and 9B variants hallucinate on 80 to 82% of knowledge tasks, per the post.

For anyone building on these models, the fix isn't a bigger download. Pair a quantized 20 to 40B reasoning model with an external retrieval layer. A wrong fact then becomes a data-source correction, not a retraining job.