Meteorite logoMeteorite
Back to videos

GLM 5.2 explained in 7 minutes...

GLM 5.2 just dropped and I had to break it down. It's an open-weights model from Z.ai that beats GPT 5.5 on long-horizon coding at roughly a sixth of the cost — but the part I find genuinely clever is the architecture trick underneath, called IndexShare. In this video I build it up from the fundamentals: how mixture-of-experts keeps a 744B model cheap to run, why a 1-million-token context is so expensive in the first place, and how reusing a single sparse-attention indexer across four layers cuts the compute 2.9x — which is what finally makes that million-token window affordable to actually run. Then I walk through the benchmarks and the economics, and I tell you why I think this efficiency-first playbook from the Chinese labs is the thing the frontier should be worried about. Sources I used: GLM 5.2 specs (vLLM): https://recipes.vllm.ai/zai-org/GLM-5.2 IndexShare + benchmarks (Latent Space): https://www.latent.space/p/ainews-glm-52-the-top-frontend-coding Benchmarks & cost (VentureBeat): https://venturebeat.com/technology/z-ais-open-weights-glm-5-2-beats-gpt-5-5-on-multiple-long-horizon-coding-benchmarks-for-1-6th-the-cost #GLM #llm #deeplearning Chapters 00:00 Intro 00:39 Same playbook, again 01:07 Mixture of Experts 01:47 The long-context wall 02:17 Sparse attention 02:59 IndexShare 03:40 Faster inference 04:35 Benchmarks 05:22 The economics 05:52 Caveats & my take

Suggested videos