Z.ai's smallest model is pulled 2.8 million times — seven times more than its whole catalogue combined

Z.ai publishes almost everything it builds on Hugging Face under an MIT licence, which makes its download counters an unusually honest ranking of what people actually run. The ranking does not look like the price list. At the top sits GLM-OCR, released at the end of January 2026: a document parser with 0.9 billion parameters, or 1.33 billion once the vision encoder in the weight files is counted. It has been downloaded roughly 2.85 million times. Behind it, the models the company leads with are barely visible — GLM-4.5-Air at 172,000 pulls, the GLM-4.5 flagship at 115,000, the GLM-4.5V vision model at 97,000, the newer GLM-4.6V at 8,400. Added together, the four come to under 400,000: about a seventh of the OCR model on its own. The reasons are not mysterious. A 0.9-billion-parameter model fits on hardware most organisations already own, deploys under vLLM or SGLang and is fast enough to sit inside a document pipeline rather than behind an API call. A 355-billion-parameter mixture-of-experts network does not fit anywhere cheap. And the job GLM-OCR does — turning scans, photographs and PDFs into text, HTML tables or structured JSON — is one that every bank, insurer, hospital and registry has, in volume, forever. The pricing tells the same story from the other end. Z.ai charges $0.03 per million tokens for GLM-OCR in either direction, the cheapest entry in its list and a twentieth of what the flagship costs to read the same text. The company reports 94.62 points on OmniDocBench V1.5, and its documentation stresses the awkward cases rather than the easy ones: stamps, dense tables, code listings, handwriting, mixed text and images. For a catalogue that tracks flagship launches month by month, it is a useful corrective. The model that gets used is not always the model that gets announced.