Sarvam

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. As adoption grew, we began serving users with considerably higher document volumes. So we reworked parts of our serving stack - optimized inference kernels for state-space architecture, smarter page-level batching, better hardware utilization across our sovereign cloud. The result is a model that runs far more efficiently at scale. We're passing these gains directly to our users through this price drop. Document intelligence should be within reach of every builder and enterprise in India. From invoices and medical records to contracts and historical archives, Sarvam Vision is now built for the scale you need. Try it: https://lnkd.in/gf9Jtv2a Docs: https://lnkd.in/gSHDjF4V

We're hosting a webinar on Saaras V3, Sarvam's Speech-to-Text model for Indian languages, to help you power your Voice AI use cases with Indian speech-to-text.We'll walk through how Saaras V3 handles real Indian speech, and what it takes to ship production-ready Voice AI with it. Here's what we'll cover: - How ASR fits into the broader Voice AI stack, across transcription, dictation, voice agents, and call analytics - What makes Saaras V3 work for Indian speech: language coverage, code-mixing, regional accents, and noisy audio - Why diarization matters in call centres and meetings, where separating speakers shapes context and intent - Streaming vs batch transcription, and the tradeoffs between latency, accuracy, and product experience - How Saaras V3 is running in production across customer support, fintech, and consumer apps.

We're hosting a webinar on Saaras V3, our speech-to-text model for teams building Voice AI for India. Date: Thursday, May 21 Time: 5:00 PM IST In this session, we'll break down what it takes to ship Voice AI that works reliably across noisy environments, regional accents, mixed-language speech, and multiple speakers. Register here: https://lnkd.in/g4qxyB2U

𝗜𝗻𝘁𝗿𝗼𝗱𝘂𝗰𝗶𝗻𝗴 𝗔𝗰𝘁𝗶𝘃𝗮𝘁𝗲 𝗔𝗜 𝗙𝗲𝗹𝗹𝗼𝘄𝘀 India’s best student AI builders are already out there. Shipping side projects at 2am, contributing to open source, fine-tuning models in dorm rooms, building things that don’t quite exist yet. We want to help them succeed. Activate AI Fellows is an 8-week summer fellowship placing 15 exceptional student builders inside India’s leading AI startups in Bangalore. Fellows work directly on real product and engineering problems while becoming part of a tight cohort of ambitious builders. 𝗪𝗵𝗼 𝘄𝗲’𝗿𝗲 𝗹𝗼𝗼𝗸𝗶𝗻𝗴 𝗳𝗼𝗿. People who ship fast, think independently, and have founder energy. We don’t care where you studied. We care about what you’ve built. 𝗪𝗵𝗮𝘁 𝘆𝗼𝘂’𝗹𝗹 𝗱𝗼. Work on real problems. Get paid. Meet exceptional founders and builders. Build relationships that compound for years. 𝗪𝗵𝗮𝘁 𝘄𝗲 𝘄𝗮𝗻𝘁 𝗯𝗮𝗰𝗸. Nothing. Just show up and build. Program starts June 1st. Applications close on 15th May. https://lnkd.in/ezFa2_ht Pratyush Choudhury ☁️ Aakrit Vaish Sriram Rajamani Garvit Juniwal Nirant Kasliwal Anshul Bhide Debdoot Mukherjee

Sarvam is partnering with YCP India to accelerate enterprise AI in India. Together, we're pairing Sarvam's full-stack sovereign AI platform with YCP's deep cross-sector expertise and execution muscle to identify the highest-impact use cases and help enterprises capture the full value of AI.

We are excited to announce that Sarvam is partnering with Pixxel to power the AI backbone of India's first orbital data centre satellite. This is a first for the country. India-built AI models will run on an India-built satellite, with both training and inference happening directly in orbit, without any dependence on foreign cloud or ground infrastructure. Most satellites collect data and ship it home for processing. This mission changes that. The demonstrator satellite will carry datacenter-class GPUs in orbit, the same hardware that powers AI on the ground, with our models running on board and analysing hyperspectral imagery in real time as it is captured. For us, this is the natural next step. AI infrastructure is a sovereignty question, and we have been building India's full-stack AI platform from the ground up. The goal has always been to make intelligence accessible to everyone, everywhere. Now, everywhere includes space. Proud to be on board this mission. Read more about the partnership here: https://lnkd.in/gCf32ZTa

At AWS Summit Bengaluru 2026, Pratyush Kumar spoke about what’s next for Sarvam, focusing on distribution at scale and making AI accessible to every citizen in India in a meaningful way. From foundation models trained from scratch to the full stack of how AI actually reaches people, the talk covered the bet we’ve taken and where we’re taking it next. Sarvam was present at the Startup Zone, and the energy from developers and students was hard to miss. The push to build cutting-edge AI in India has never been stronger. We’re backing builders through our Startup Program with API credits, ready-to-use infrastructure, and hands-on support. The goal is to activate the ecosystem and empower builders to build with AI. Apply to the program: https://lnkd.in/gva7jrGb

We're open-sourcing two frameworks for evaluating Indian ASR, and a full guide on evaluation across 22 languages. WER (Word Error Rate) and CER (Character Error Rate) were built for languages like English. Indian languages behave differently. Formal and colloquial forms coexist. English words move fluidly between scripts. In this guide, we introduce four complementary metrics LLM-WER, LLM-CER, Intent Score, and Entity Preservation Score. In combination they capture a more complete view of Indian ASR system performance. Read more: https://lnkd.in/gPd5CWsi Both frameworks are available to use and build on for evaluating Indian ASR in real-world settings. Open-source frameworks: LLM-WER/CER: https://lnkd.in/g-h3jn6n Intent + Entity Score: https://lnkd.in/g4GBkhg8

At GitHub Constellation, the Sarvam engineering team shared their journey scaling large language models from early pre-training to post-training a 105B Mixture-of-Experts model. In the session, they covered key challenges across large-scale pretraining, reinforcement learning, and multimodal capabilities across vision and speech, with a focus on real engineering trade-offs and lessons learned.