GLM-OCR

Z.ai's compact document model, backfilled here: a 0.9B multimodal model pairing a 0.4B CogViT visual encoder with a 0.5B GLM language decoder for real-world document understanding. It adds multi-token prediction to escape the inefficiency of autoregressive decoding on deterministic OCR, predicting several tokens per step with shared parameters, and runs as a two-stage system in which PP-DocLayout-V3 performs layout analysis before recognition. The March 2026 report puts it first on OmniDocBench V1.5 at 94.62. MIT licence; the repository has passed 1.7 million downloads.