Surya OCR version 3!

The latest version of Surya OCR has a new architecture, and is trained on significantly more data than before.

Some notable features:

• 90+ language support
• Handles inline math and equations
• Char, word, and line bboxes available
• Significantly better chinese performance
• Very fast - 10000+ tokens/second on A100 with vllm
• Continuous batching in base configuration for ~2x speedup

Updated benchmarks coming soon.

Examples

Word boxes

Math

Chinese

What's Changed

• Foundation new processor by @VikParuchuri in #342
• WIP: Continuous Batching + Other Optimizations for new model by @tarun-menta in #337
• Foundation vik dev by @VikParuchuri in #343
• Foundation vik dev by @VikParuchuri in #348
• Foundation vik dev by @VikParuchuri in #349
• Foundation vik dev by @VikParuchuri in #351
• Vik perf by @VikParuchuri in #353
• Vik perf by @VikParuchuri in #355
• Foundation flash by @tarun-menta in #354
• Vik perf by @VikParuchuri in #356
• Vik perf by @VikParuchuri in #358
• Qwen encoder by @VikParuchuri in #362
• Foundation by @VikParuchuri in #363
• Vik dev by @VikParuchuri in #365
• Surya OCR 3 by @VikParuchuri in #364

Full Changelog: v0.13.1...v0.14.0