# OCR improvements, Gemini 3.5 benchmarks, STT word level accuracy, caching and more

URL: https://interfaze.ai/blog/28-06-26-updates

Here's what we shipped this week.

## High density PDF and image resolution handling

![High density PDF and image resolution handling](https://r2public.jigsawstack.com/interfaze/blogs/28-06-26-updates/ocr_handwriting_dense.png)

High DPI PDFs and images can easily overload GPUs having to load them into memory.

We've fixed this by smart normalization preprocessing documents and images to a standardize scale without dropping performance and now with the ability to handle larger files.

Check it out: https://interfaze.ai/docs/vision/ocr

## Page selection on OCR documents

You can now easily prompt the pages you want to process for OCR tasks even if you pass in a large 50 page PDF.

## Gemini 3.5 benchmarks added

![Gemini 3.5 benchmarks added](https://r2public.jigsawstack.com/interfaze/blogs/28-06-26-updates/benchmark_28th_may_26.png)

We've added a newly released flash series model Gemini 3.5 flash to the benchmarks.

While it's an improvement from Gemini 3 flash, it's 3x the cost leaning closer to Pro tier models.

Interfaze continues to take the lead across all 7 benchmarks.

Check out the full benchmarks: https://interfaze.ai/leaderboards

## STT word level time stamps

![STT word level time stamps](https://r2public.jigsawstack.com/interfaze/blogs/28-06-26-updates/stt_word_level_example.png)

Prompt the model asking for word level timestamps for higher granularity transcriptions.

Check it out here: https://interfaze.ai/docs/audio/speech-to-text#word-level-timestamp#word-level-timestamp

## Improve file upload speed

Binary/Base64 files now upload faster on API calls.

Check out all the different file handling methods: https://interfaze.ai/docs/handling-files

## Improved caching for faster response

Caching has been improved significantly on pre-processing tasks.

For example running the same image twice once for OCR and another for object detection, the image would need to be processed twice, either compressed or normalized.

Now a big part of pre-processing is cached making it faster and slightly cheaper on multiple runs.

## Improved audio language detection

Audio language detection now has higher accuracy based on accents and other tonality attributes.

Try it with STT: https://interfaze.ai/docs/audio/speech-to-text

### Coming soon:

- Better deep search capabilities
- Improved GUI detection speed
- Improved Object detection granularity

That's it for this week!

Best,
Yoeven
CEO, interfaze.ai