Gemini 3.7 Flash vs GLM-OCR

Gemini 3.7 Flash is a proprietary multimodal model from Google, positioned in the Flash branch of the Gemini 3 series that trades some of the capacity of the larger Pro models for lower latency and lower cost per token. It accepts interleaved text and image input alongside other modalities handled by the Gemini family and returns text, and it continues the series pattern of exposing a configurable thinking budget so that reasoning effort can be scaled up for harder problems or reduced for high throughput extraction, routing and classification work. The model is announced roughly three weeks after Gemini 3.6 Flash, part of an unusually fast iteration cadence within the Flash line.

Google reports gains concentrated in agentic coding and front end generation, citing a WebDev Arena Elo of 1588 for this release compared with 1538 for the preceding Flash model, and describes it as producing more functional layouts and more feature complete applications in fewer prompts. Weights are not published and the architecture, parameter count and training corpus are undisclosed, consistent with prior Gemini releases. Visual capability follows the Flash lineage, covering image and document understanding, chart and diagram interpretation, text recognition in images, and general visual question answering.

GLM-OCR is a multimodal OCR model for complex document understanding, built on the GLM-V encoder-decoder architecture by Zhipu AI. The model combines a 0.4B-parameter CogViT visual encoder pre-trained on large-scale image-text data, a lightweight cross-modal connector with efficient token downsampling, and a 0.5B-parameter GLM language decoder, totaling 0.9B parameters. To address the inefficiency of standard autoregressive decoding in deterministic OCR tasks, GLM-OCR introduces a Multi-Token Prediction (MTP) mechanism that predicts multiple tokens per step, significantly improving decoding throughput while keeping memory overhead low through shared parameters. Training proceeds through four stages: visual encoder pretraining with MIM, CLIP, and distillation objectives; vision-language pretraining on document parsing, grounding, and VQA data; supervised fine-tuning on curated OCR datasets covering text, formula, table, and key information extraction; and full-task reinforcement learning to improve accuracy and structural consistency.

At the system level, GLM-OCR adopts a two-stage pipeline in which PP-DocLayout-V3 first performs layout analysis, followed by parallel region-level recognition. This design enables robust handling of diverse document layouts including tables, formulas, and multi-column text. The model supports document parsing and targeted recognition tasks, producing structured outputs in Markdown, JSON, and LaTeX formats across more than 100 languages. On the OmniDocBench V1.5 benchmark, GLM-OCR scores 94.62, and achieves 94.0 on OCRBench and 96.5 on UniMERNet for formula recognition.