Trillion Labs Releases GravityOCR, an OCR Model Applying Self-Speculative Decoding

Trillion Labs has released GravityOCR, an OCR model that converts document images into text, HTML tables, and LaTeX formulas. This model reuses the weights of GLM-OCR (0.9B) and applies the Self-Speculative Decoding technique to significantly improve decoding speed.

Key Technologies and Performance

GravityOCR adopts a structure that drafts via a Diffusion path and verifies via an Autoregressive (AR) path by adding only one learnable embedding for the mask token [M] to GLM-OCR, without a separate draft model or auxiliary head. This allows the model to finalize an average of 9.7 tokens per forward pass.

• Speed Improvement: Processing 0.730 pages per second on an H100 GPU, which is 1.32x faster than the AR path of the same checkpoint (0.554 pages).
• Accuracy Retention: Recorded an Overall score of 95.16 on the OmniDocBench v1.6 benchmark, maintaining 99.7% of the accuracy of the original GLM-OCR (95.48).
• Output Identity: Generates identical output to AR greedy decoding in exact arithmetic operations, showing a 96.6% exact match rate even in bf16 serving environments.

Training and Implementation Features

It was trained via fine-tuning from the GLM-OCR checkpoint, jointly learning Clean Stream (AR loss) and Corrupted Streams (Diffusion loss). The training data used approximately 12.3 million region-level examples (primarily in English). During inference, it finalizes a minimum of 2 to a maximum of B+2 tokens per round through two forward passes (draft generation and verification).

License and Public Information

Model weights are released on Hugging Face under the MIT License, and serving and inference code are available on GitHub. It can be served by applying a patch to SGLang v0.5.12, with the --spec_natcache option ensuring output identical to the AR path at the byte level. However, the training code and data pipeline remain private.