GM @SarvamAI | previously microsoft research, everwell, harvard university
Joined February 2025
- Tweets144
- Following390
- Followers658
- Likes225
Based on requests, I’m sharing the slide deck from my talk on training **Sarvam Vision 1.0** at @aiDotEngineer World's Fair! In the presentation, I discussed how we built India's first sovereign 3B state-space vision model to solve OCR for English + 22 official Indian languages.
As a follow up, we recently released **Sarvam Vision 2.1**. It delivers a step change in capabilities for document AI. 2.1 adds **Key-Value form extraction, complex multi-page table parsing, and Indic handwritten recognition**. Additionally, we eliminated model hallucinations and optimized inference costs.
youtube.com/watch?v=T72nqdC9…
In 2.1, we also rethought how Doc AI benchmarks evaluate various capabilites that do not necessarily correlate. This required framing evaluation as a Pareto optimization problem across benchmarks whose metrics rarely align. We mapped this across two distinct frontiers: global (olmOCR-Bench vs. OmniDocBench) and language breadth (English vs. Indic). In both, Vision 2.1 pushes the Pareto frontier, unlocking new regions.
Read more in the blog I have written: sarvam.ai/blogs/sarvam-visio…
Slides: drive.google.com/file/d/1MzY…
To follow up, we recently released **Sarvam Vision 2.1**. It delivers a step change in capabilities for document AI. 2.1 adds **Key-Value form extraction, complex multi-page table parsing, and Indic handwritten recognition**. Additionally, we eliminated model hallucinations and optimized inference costs.
In 2.1, we also rethought how Doc AI benchmarks evaluate various capabilites that do not necessarily correlate. This required framing evaluation as a Pareto optimization problem across benchmarks whose metrics rarely align. We mapped this across two distinct frontiers: global (olmOCR-Bench vs. OmniDocBench) and language breadth (English vs. Indic). In both, Vision 2.1 pushes the Pareto frontier, unlocking new regions.
Read more about 2.1 in the blog I have written: sarvam.ai/blogs/sarvam-visio…
krishna retweeted
Well under 1% of the common corpus used to train frontier language models has Indian-language representation, according to a published language distribution study. @fewshotlearner, Head of Vision Models at Sarvam, takes on that gap in "From Scratch to SOTA: Training a 3B State-Space Vision Model," which @aiDotEngineer published on YouTube.
The talk walks through how a 3 billion parameter model, small enough to run on a single GPU, got to state of the art in document OCR and beat models 100 times its size, with the data, training, and compute all done in India.
- State-space, not transformer. Krishna lays out the tradeoff: a transformer's attention cost grows with the square of sequence length, while a state-space model updates a single running state token by token, so compute scales linearly and memory stays constant, which matters when a page can produce 5,000 to 10,000 visual tokens.
- A contrarian bet on block-level OCR. Where most VLMs in late 2025 did page-level OCR, Sarvam Vision does OCR at the block level instead, with two dedicated modules, one for layout and one for reading.
- A four-stage curriculum. Text-only pretraining on 13 trillion tokens across English, Indian languages, math, and code builds the language backbone first. Then 300 million image-text pairs teach it to see, 100 million OCR samples take it through supervised fine-tuning, and a reinforcement learning stage (sample a group, score against unit tests, reinforce toward the average to beat) pushes past what fine-tuning alone reaches.
- The numbers. At launch it scored 84.3 on the Olmo OCR Bench and 93.2 on Omnidoc Bench, and it holds an unbeaten lead across 22 Indian languages, ahead of frontier models including Gemini, ChatGPT, and Opus.
- Already in production. The model powers Akshar, an agentic document intelligence workbench with confidence scores, block-level grounding, and agentic proofreading, and has been used to digitize more than 35 million pages across English and 22 Indian languages.
I'm working through the published talks from AI Engineer World's Fair sharing summaries and takeaways. Follow for more!
krishna retweeted
What an inspiring talk by @fewshotlearner from @SarvamAI at @aiDotEngineer.
From Scratch to SOTA: Training a 3B State-Space Vision Model.
youtube.com/watch?v=T72nqdC9…
“Well under one percent of the Common Crawl corpus that frontier models train on is in an Indian language, even as frontier labs call India a fast growing market. Krishna Prasad Srinivasan says the knowledge exists but was never digitized, and Sarvam's answer is a three billion parameter vision language model, small enough for one GPU, that he says beats document AI models a hundred times larger. The model is unusual twice over. Its backbone is a state space model rather than a transformer, because a page can run to ten thousand visual tokens and quadratic attention gets expensive, while an SSM keeps constant memory. And in late 2025, when most OCR models were monolithic page level VLMs, the team bet on block level OCR wrapped in a layout harness and a reading order harness, which many 2026 releases have since converged on.”
For my last hurrah at Sarvam: super excited to launch Sarvam Vision 2.1! It's SOTA on olmOCR-Bench (87.3) and on our Indic OCR benchmark (87.39). With new capabilities like structured extraction and Indic handwriting recognition, it's a step change for document intelligence workflows 🧵
Sarvam Vision 2.1 is live, with new capabilities for structured extraction and Indic handwritten recognition across English and 22 Indian languages.
The model is state of the art on both olmOCR-Bench and our own Indic benchmark, scoring 87.3 and 87.39 respectively.
Read more in the blog: links.sarvam.io/zF7SiK
To support more rigorous benchmarking of Indian languages, we release the Sarvam Indic OCR Bench. It has 6,909 samples in 22 Indian languages and English, drawn from newspapers, textbooks, brochures and historical writing dated from 1800 to today.
The benchmark contains samples with varying difficulty to measure OCR accuracy.
Link: huggingface.co/datasets/sarv…
Try Vision 2.1 now!
- Digitize API: docs.sarvam.ai/api-reference…
- Extract API: docs.sarvam.ai/api-reference…
- Akshar Document Agents: indus.sarvam.ai/akshar/dashb…
the talks from the vision track at AI Engineer WF 2026 are out! as part of an exciting line up showcasing cutting-edge work, i presented the Sarvam Vision 1.0 model
Live now: our Vision & OCR Track from AI Engineer World's Fair 2026.
A model that counts 32 white squares on part of a chessboard. A file format that stores a table as a pile of line segments. Ten turkeys on the roof of a Tesla.
Thesis: the models can see. They are still learning to look.
youtube.com/watch?v=RQi7x-na…
- Building the Document Context Layer for AI Agents: @jerryjliu0, LlamaIndex
- Skill issue: stop deploying vision language models, use them with Skills: @mervenoyann, Hugging Face
- Modality Misalignment and Originality Attribution in Short-Form Video: Aditya Gautam, Meta
- From Ingestion to Agents: How AI Teams Build on Document Intelligence: Adit Abraham, Reducto
- The Best Models Still Reason Like Toddlers: @andrewdai, Elorian
- You're Not Thinking Big Enough: Rebuilding Food Systems with AI Agents: @cbmenefee, Firecrawl
- From VLM/VLA's to Embodied Agents: @ArmenAgha, Perceptron AI
- From Scratch to SOTA: Training a 3B State-Space Vision Model: @fewshotlearner, Sarvam
Excited to present our work on Sarvam Vision at @aiDotEngineer World's Fair in SF. Over the last year, we have built India's first SOTA sovereign VLM for document intelligence. In my talk, I discuss what we did and how we did it.
See you at Room 2006, Moscone Center. Happy to chat more about all the other work we are doing after the talk!
@SarvamAI @SarvamForDevs @pratykumar
krishna retweeted
Sarvam joined the AI discussions at the 52nd G7 Summit in Évian-les-Bains, France.
@vivekrag participated alongside G7 leaders and chief executives from some of the world’s leading AI companies. The discussion centered on frontier AI risks, infrastructure, and sovereignty in the age of AI.
A significant moment for Sarvam and for India’s AI ecosystem.
krishna retweeted
We're thrilled to announce that we have raised $234M in the first close of our $300M Series B at a $1.5B valuation.
@HCLTech and @BessemerVP have joined us in this round, alongside continued support from @khoslaventures and @peakxvpartners
For countries and companies, sovereign control on the AI stack is no longer an optionality. Sarvam will be the partner of choice for this aspiration. The capital allows us to accelerate our momentum towards this full stack of models, compute, and deployments.
A huge thank you to our customers, partners, investors, and the Sarvam team for your trust and belief in what we are building. We’re just getting started.
Read more: sarvam.ai/announcing-series-…
krishna retweeted
Unlocking document intelligence for India scale efficiently!
Earlier this February, we launched Sarvam Vision, a vision-language model for document intelligence.
Today, more than 35 million pages are being digitised through the Sarvam Vision API by developers and partners.
Since launch, we've made it significantly more efficient to serve at scale. We’re now passing these gains on by reducing the Sarvam Vision API price from ₹1.5 to ₹0.5 per page.
Sarvam Vision, our SOTA document intelligence model, is now 66% cheaper!
since launch, thousands of developers have digitized millions of complex documents with it. we've optimized our serving stack to scale with that demand - and we're passing on all of the gains to users.
excited to see all that you build with the model. @SarvamAI @SarvamForDevs
Earlier this February, we launched Sarvam Vision, a vision-language model for document intelligence.
Today, more than 35 million pages are being digitised through the Sarvam Vision API by developers and partners.
Since launch, we've made it significantly more efficient to serve at scale. We’re now passing these gains on by reducing the Sarvam Vision API price from ₹1.5 to ₹0.5 per page.