Souvik Mandal

Deep Learning Engineer @ Nanonets. I train VLMs for document understanding

[Jun Song @jun_song]

Opus 4.7 토큰 테스트

토크나이저 차이로 제미나이의 2배를 사용합니다.
Opus 4.6 대비해서도 50% 많이 사용해요.

이건 사실상 같은 한도에서 모델이 50% 더 비싸진겁니다.

[Souvik Mandal @mandalsouvik4]

@giffmana 🤔🤔

[Souvik Mandal @mandalsouvik4]

@MaziyarPanahi @Yuchenj_UW @sama Whoever forked the repo now should apply for claude for open source now, more than 5k stars 😂

[Souvik Mandal @mandalsouvik4]

Maybe Claude code open sourced its codebase because it cannot figure out all the bugs related to high token usage and open sourcing it is more efficient. 🚀

[Rushabh Nagda @rushabh_nagda]

"When a metric becomes the target, it stops being a good metric" - Goodharts Law

last few days GLM-OCR has been trending after it claimed 95% on OmniDocBench, which is higher than Gemini-3-pro

in reality GLM-OCR is way worse than the story these benchmarks paint, lets see how

full disclaimer: ive been working in this space for the last 7 years with @nanonets

[Y Combinator @ycombinator]

Every student accepted into Startup School India now gets $25k+ in AI and cloud credits.

Apply, get in, and start building: https://t.co/gncXSJGhdb

[Souvik Mandal @mandalsouvik4]

@vanstriendaniel Nanonets OCR 3!! Preview model is available for test at https://t.co/oQfNvwx1WT

[Souvik Mandal @mandalsouvik4]

@techNmak Model is not available on hf. Some issue with model configuration https://t.co/JyDMO7CYdF

[Souvik Mandal @mandalsouvik4]

@vanstriendaniel @huggingface Results of Nanonets OCR 3 intermediate checkpoint works fine. https://t.co/gIettdM1Aj. We will open-source it soon!!

[Souvik Mandal @mandalsouvik4]

@ihailmyindia Models are not released yet.

[Rann @rstone_9]

A tiny phishing trick broke 5 out of 8 popular OCR models (Deepseek-OCR, Gemini-2.5Pro, PaddleOCR, etc)

Can you spot the phish in the screenshot?

If you use OCR/Markdown pipelines in prod, check out the results in the thread.

1/n

[Maziyar PANAHI @MaziyarPanahi]

> RTX Pro 6000 96GB @PrimeIntellect
> Nanonets-OCR @nanonets
> vLLM @vllm_project
> 2000 images from cifar10 @huggingface

one word, damn!

[Souvik Mandal @mandalsouvik4]

@HKydlicek Feel free to try the newer model. We have fixed some of the issues of the last model (Nanonets-OCR-s) that people mentioned. Also, added some of the requested functionalities.

[Souvik Mandal @mandalsouvik4]

Introducing Nanonets-OCR2: a lightweight 3B VLM that transforms documents into clean, structured Markdown and is capable of VQA. We have trained the model on close to 3 million documents. It is multi-lingual and can handle handwritten documents.

[Vaibhav (VB) Srivastav @reach_vb]

Nanonets released a new version of their SoTA OCR model 🔥

Supports LaTeX, Multilingual, Complex tables and much more

Works out of the box with transformers, vllm and all major runners! 🤗

https://t.co/fgvdp4QIC9

[Souvik Mandal @mandalsouvik4]

@BramVanroy @vanstriendaniel @nanonets @huggingface We have compared against Gemini-Flash, which is generally the SOTA for document understanding tasks. https://t.co/bqZLJVGxt9 We will add more evals more time.