Sarvam AI

Sarvam AI, which brands itself simply as Sarvam and is legally Axonwise Private Limited, is a Bengaluru-based artificial intelligence company that builds large language models, speech, translation and document models for Indian languages, together with the products and serving infrastructure to run them inside India.^[1][2] It was founded in August 2023 by Vivek Raghavan and Pratyush Kumar, who had both worked on the AI4Bharat initiative at the Indian Institute of Technology Madras, and it describes itself as "India's full-stack sovereign AI company".^[2][3][4] Kumar is the company's chief executive.^[5]

In April 2025 the Government of India selected Sarvam under the IndiaAI Mission to build a sovereign foundation model, with subsidised access to GPU compute.^[6][7][8] The company unveiled Sarvam 30B and Sarvam 105B, two mixture-of-experts reasoning models that it says were trained from scratch in India on compute provided under the mission, at the India AI Impact Summit 2026 in February 2026, and released their weights under the Apache License 2.0 on 6 March 2026.^[9][10] In June 2026 it announced a $234 million first close of a $300 million Series B round at a post-money valuation of $1.5 billion, with the IT services company HCLTech committing $150 million as lead strategic investor.^[4][11] Entrackr reported in August 2026 that its board had approved a further $74 million extension led by Nvidia, and The Economic Times reported in October 2026 that it was in talks to raise $250 million to $300 million at a valuation of about $3 billion.^[12][13]

Overview

The word sarvam means "all" in Sanskrit; Kumar said in 2023 that the name reflected the company's aim to make AI "accessible to all".^[15]

History

Founding and first funding (2023)

Sarvam came out of stealth in December 2023, when TechCrunch reported that the five-month-old startup had raised $41 million across its seed and Series A rounds.^[3] Lightspeed Venture Partners led the Series A and co-led the seed round with Peak XV Partners, and Peak XV and Khosla Ventures also took part in the Series A.^[3] The company's own announcement described the raise as a $41 million Series A led by Lightspeed and called it the largest raise at that stage for an Indian AI startup.^[15] At the time Sarvam had about 18 employees and planned to build language models for Indian languages that used voice as the default interface, along with a platform for businesses to build and deploy applications on those models.^[3]

About a week after the funding announcement, Sarvam released its first model, OpenHathi-Hi-v0.1.^[16] It was a 7-billion-parameter base model built on Meta's Llama 2 and trained further on Hindi, English and Hinglish, published on Hugging Face under the Llama 2 license and developed with AI4Bharat as an academic partner.^[17][16]

Platform launch and Sarvam-1 (2024)

On 13 August 2024 Sarvam launched what it called a full-stack generative AI platform supporting 10 Indian languages.^[18] The launch included Sarvam Agents, voice-enabled business agents that could be deployed over the telephone, WhatsApp or inside apps with prices starting at 1 rupee per minute; Sarvam 2B, a 2-billion-parameter model the company said was trained from scratch on an internal dataset of 4 trillion tokens; Shuka 1.0, an audio language model; APIs for translation, speech recognition, speech synthesis and document parsing; and A1, a generative AI workbench for lawyers.^[18] Shuka v1 joined Sarvam's in-house Saaras v1 audio encoder to Meta's Llama 3 8B Instruct through a small projector of about 60 million parameters, and only the projector was trained, on less than 100 hours of audio.^[19] Later that month the company released Bulbul v1, a code-mixed, multilingual text-to-speech model.^[20]

On 24 October 2024 Sarvam released Sarvam-1, a 2-billion-parameter model for 10 Indian languages and English.^[21] The company built a training corpus of about 2 trillion Indic tokens, which it called Sarvam-2T, added roughly equal amounts of English and a body of source code, and designed a tokenizer that needs 1.4 to 2.1 tokens per word across the supported languages.^[21][22] The model was trained on Yotta's Shakti cluster on 1,024 GPUs over five days using NVIDIA NeMo, and it was released under a Sarvam non-commercial license.^[22]

UIDAI deployment and IndiaAI Mission selection (2025)

On 18 March 2025 the Unique Identification Authority of India (UIDAI), which runs the Aadhaar identity system, announced a partnership under which Sarvam would deploy voice-based AI for Aadhaar holders.^[23] The system collects feedback on enrolment and update services, including reports of overcharging, and sends fraud alerts when it spots suspicious authentication requests, in 10 languages at launch.^[23] According to the government release, Sarvam delivered a custom generative AI stack hosted on-premises inside air-gapped UIDAI infrastructure, the agreement ran for one year with an option to extend by another, the work was done by volunteers from Sarvam with UIDAI's Technology Centre in Bengaluru under UIDAI's volunteer policy, and UIDAI owns the solution.^[23] Sarvam later said the deployment also handles digital life certificates for pensioners and age verification for residents over 70, across 4 to 6 crore (40 to 60 million) residents.^[24]

On 26 April 2025 the IndiaAI Mission selected Sarvam to build what the government called India's Sovereign LLM Ecosystem, "developing an open source 120 billion parameter AI model" for governance and public-service uses.^[6] The same release said the mission's foundation model programme had received 506 proposals by 30 April 2025, and announced three further selections: Soket AI, Gnani AI and Gan AI.^[6] Kumar said Sarvam would develop three variants, Sarvam-Large, Sarvam-Small and Sarvam-Edge, in collaboration with AI4Bharat at IIT Madras.^[7] In May 2025 the company said it would open-source the models it trained under the mission under permissive licenses, and named AI4Bharat and IIT Madras's Centre for Responsible AI (CeRAI) as close collaborators.^[25]

Inc42 described Sarvam as one of the largest beneficiaries of the scheme's compute allocation, reporting in June 2026 that it received a subsidy of 98.68 crore rupees against a total compute bill of 246.71 crore rupees, covering access to 4,096 Nvidia H100 GPUs for six months, with the government paying 40% of the cost.^[8] An NVIDIA case study says Sarvam achieved near-linear training scaling across 4,096 or more H100 GPUs, running on Yotta's cloud with NVIDIA Quantum InfiniBand networking and Megatron-LM parallelism.^[26]

Sarvam-M and Sarvam-Translate (2025)

In May 2025 Sarvam released Sarvam-M, a 24-billion-parameter hybrid reasoning model with "think" and "non-think" modes.^[27][28] Unlike Sarvam-1, it was not trained from scratch: the company applied supervised fine-tuning and reinforcement learning with verifiable rewards, using GRPO, to Mistral AI's Apache-licensed Mistral Small, and said the "M" stands for Mistral.^[27] Sarvam reported average gains over the base model of 20% on Indian-language benchmarks, 21.6% on math and 17.6% on programming.^[27] The release drew public criticism over its low download count and its foreign base model (see Reception and criticism below).

In June 2025 the company released Sarvam-Translate, a document-level translation model for 22 Indian languages built with AI4Bharat as a fine-tune of Google's Gemma 3 4B instruction-tuned model; its Hugging Face repository lists the GPL-3.0 license.^[29]

Sovereign model launches (February to March 2026)

In February 2026 Sarvam ran what Business Today called a 14-day launch streak.^[30] According to its blog, it released Sarvam Dub (1 February), Sarvam Audio (3 February), the Sarvam Vision document model (5 February), the Bulbul V3 text-to-speech model (7 February), the Sarvam Arya agent orchestration stack (10 February), the Saaras V3 speech recognition model (11 February), Sarvam Studio (13 February), Sarvam Edge for on-device models (14 February) and the Akshar document workbench (15 February).^[31]

On 9 February 2026 the company announced partnerships with two state governments.^[32] Odisha signed a memorandum of understanding with Sarvam on 6 February for a sovereign AI capacity hub, for which Sarvam said it was planning a 50 MW AI-optimised facility. In Tamil Nadu, Sarvam and IIT Madras are to build "Digital Sangam", which the company described as India's first Sovereign AI Research Park, around a 20 MW AI-optimised data center.^[32]

At the India AI Impact Summit in New Delhi, Sarvam unveiled the 30-billion and 105-billion-parameter models together with text-to-speech, speech-to-text and vision models.^[9] It said the models were trained using compute provided under the IndiaAI Mission, with infrastructure from the data center operator Yotta and technical support from Nvidia.^[9] At the same summit Sarvam announced a partnership with HMD to put an AI assistant on Nokia and HMD feature phones, said it had tuned its edge models for Qualcomm chipsets and was working with Bosch on in-car assistants, and showed Sarvam Kaze, AI smart glasses designed and manufactured in India.^[33]

On 20 February 2026 the company opened Indus, a chat app for web, iOS and Android powered by Sarvam 105B, in a limited beta with a waitlist because of constrained compute.^[34][35] In May 2026 HMD launched the Vibe 2 5G smartphone with the Indus app preloaded.^[36] The 30B and 105B weights were published on Hugging Face and the government's AI Kosh platform on 6 March 2026, nearly three weeks after the summit announcement.^[10][37] On 16 March 2026 NVIDIA named Sarvam as one of eight inaugural members of its Nemotron Coalition for building open frontier models, alongside Mistral AI, Thinking Machines Lab and others.^[5]

Series B, Epoch and expansion (2026)

Reports of a large funding round circulated through the spring. In March 2026, Storyboard18, citing Moneycontrol, reported that Sarvam was in talks to raise $200 million to $250 million from Nvidia, Accel and HCLTech at a valuation of about $1.5 billion.^[38] On 2 April Bloomberg reported that Sarvam was close to raising $300 million to $350 million at a valuation of $1.5 billion to $1.55 billion, with Bessemer Venture Partners expected to lead and Nvidia, Amazon and Prosperity7 Ventures also expected to participate.^[39]

Sarvam announced the round on 15 June 2026: it raised $234 million in the first close of a $300 million Series B at a post-money valuation of $1.5 billion; HCLTech committed $150 million as the lead strategic investor, Bessemer invested, and Khosla Ventures and Peak XV continued their support.^[4] HCLTech rather than Bessemer was therefore the lead investor, and Nvidia, Accel and Amazon were not among the investors named in the announcement.^[4][11] TechCrunch described Sarvam as India's newest AI unicorn.^[11] Inc42 reported that HCLTech's investment bought a 10.46% stake.^[8] Sarvam said the money would fund research on its next frontier model for agentic, coding and cybersecurity uses and give it access to compute at scale.^[4]

Sarvam held its first developer conference, Sarvam Epoch, in Bengaluru in late July 2026.^[14][40] Kumar told the audience that the company was "building a trillion-plus parameter model right here in India", trained from scratch for coding, cybersecurity, simulation and science.^[14] Sarvam also launched Sarvam Inference, an India-hosted service that serves Sarvam 105B alongside open-weight models such as GLM 5.2 and Gemma 4; announced an office in San Francisco; and appointed as an advisor Devendra Singh Chaplot, whom Inc42 described as part of the founding teams at Mistral AI and Thinking Machines Lab.^[14] In its own summary, Sarvam said it released updated 105B variants for conversational and work agents, pricing the conversational variant at $0.80 per million blended tokens, made its Voice Agents product generally available at 3.50 rupees per minute, put a coding agent called Sarvam Code into early beta, previewed a security agent called Sarvam Cyber, released a Kivi voice app for Mac, and said Sarvam Inference ran on "the first fleet of Blackwell GPUs in India".^[40]

In early August, Entrackr reported from a Registrar of Companies filing that Sarvam's board had approved a $74 million (about 698 crore rupees) extension of the Series B, led by Nvidia with about $25 million, followed by Glade Brook Capital with about $20 million, and including Gaja Capital, Indigo Ventures and several individual investors and family offices.^[12] By Entrackr's analysis the extension valued Sarvam at about $1.51 billion post-money and, once allotted, would leave Nvidia with 1.66%, HCLTech with 9.95% and the founders with a combined 38.44%.^[12] On 5 August the Free Press Journal, citing The Economic Times, reported that Kumar had confirmed the extension and said that "Sarvam is the first Indian company to receive an investment from Nvidia".^[41] The same day Sarvam launched Sarvam Circle, a partner programme for cloud providers, hardware makers, system integrators and consultancies, with statements from Amazon Web Services, HCLTech, Thoughtworks, IBM, HP and YCP.^[42]

The company continued to update its speech and vision models through the autumn, releasing Sarvam Vision 2.1 on 24 September, Saaras V4 on 25 September and Bulbul V4 Flash on 13 October 2026.^[43][44][45] On 13 October 2026 The Economic Times reported, citing people aware of the discussions, that Sarvam was in talks to raise another $250 million to $300 million at a valuation of $3 billion, double its June valuation.^[13] CXO Digitalpulse noted that the funding remained unconfirmed.^[46]

Founders

According to a Storyboard18 profile citing his LinkedIn page, Raghavan spent nearly two decades in electronic design automation, holding senior roles at Avant!, Synopsys and Magma Design Automation, and founded two EDA companies.^[48] TechCrunch reported in 2023 that he had spent more than a decade at UIDAI.^[3] Sarvam's 2023 funding announcement described him as reflecting on "his decade long experience in building digital public infrastructure including Aadhaar".^[15]

Kumar's doctoral thesis at ETH Zurich, completed in 2014, was titled "Hard Real-Time Guarantees in Cyber-Physical Systems".^[47] At IIT Madras he co-founded One Fourth Labs, which ran affordable online AI courses, and started AI4Bharat as a working group on practical AI problems.^[47] Sarvam's 2023 announcement described AI4Bharat as a research initiative on open-source Indian-language AI.^[15] Both founders appear as co-authors on AI4Bharat papers, including Samanantar (2021), a collection of parallel corpora for 11 Indic languages, and IndicTrans2 (2023), translation models covering all 22 scheduled Indian languages.^[49][50] TechCrunch has described AI4Bharat as an initiative backed by the technology entrepreneur Nandan Nilekani.^[11]

Models

Sarvam 30B and 105B

Both models are sparse mixture-of-experts Transformers that use 128 experts in their expert layers.^[10] Sarvam 30B uses grouped-query attention, while Sarvam 105B uses Multi-head Latent Attention, a compressed attention scheme that reduces key-value cache memory for long inputs.^[10] The 105B model routes each token to 8 of its 128 experts plus one shared expert and supports a 128,000-token context through YaRN scaling; the 30B model has 19 layers and routes to 6 experts.^[52][51] Sarvam told TechCrunch that the 30B model supports a 32,000-token context window for real-time conversation.^[9]

Sarvam's technical post says the 30B model was pre-trained on 16 trillion tokens and the 105B model on 12 trillion, in three phases that ended with a long-context extension, using sigmoid-based routing scores with an expert-bias term to balance load.^[10] Supervised fine-tuning used synthetic and curated prompts and agentic traces from simulated environments and real repositories, and a safety dataset covering standard and India-specific risk scenarios.^[10] The reinforcement learning stage used an asynchronous GRPO system without a KL-divergence penalty and a policy objective inspired by CISPO.^[10] The tokenizer covers all 22 scheduled Indian languages across 12 scripts.^[10] Sarvam 30B powers the company's Samvaad conversational agent platform, and Sarvam 105B powers Indus.^[10]

The company reported the following scores for Sarvam 105B against open models of similar size. They are self-reported and had not been independently verified when the weights were released (see Reception and criticism below).^[10][37]

On SWE-bench Verified, Sarvam reported 45.0 for the 105B model, below GLM-4.5-Air (57.6) and gpt-oss-120b (50.6).^[10] For Indian languages the company built its own benchmark of 110 English prompts translated into the 22 scheduled languages in both native and romanised script, scored by pairwise comparison with Gemini 3 as the judge; it reported that Sarvam 105B won on average 90% of the pairwise comparisons across the judged dimensions.^[10]

Speech, translation and vision models

Sarvam's speech recognition models are called Saaras. Saaras V3, released in February 2026, added streaming transcription and covers 22 Indian languages and English; Sarvam said it was trained on more than 1 million hours of audio and reached a word error rate of about 19% on the IndicVoices benchmark, down from about 22% for Saaras V2.5.^[54] Saaras V4 pairs an audio encoder with a 3-billion-parameter hybrid state-space language model decoder trained from scratch, and offers five output formats (transcribe, translate, verbatim, transliteration and code-mixed).^[44]

The Bulbul models generate speech. Bulbul V3 offered more than 35 voices in 11 Indian languages. Sarvam said that in a blind A/B listening study run by Josh Talks against models including ElevenLabs v3 alpha and v2.5 Flash and Cartesia Sonic-3, Bulbul V3 was the most-preferred model for naturalness on 8 kHz telephony audio.^[55] Bulbul V4 was shown as a research preview at Epoch, and Bulbul V4 Flash, a faster version for real-time use, followed in October 2026.^[40][45]

Sarvam Vision is a 3-billion-parameter state-space vision-language model aimed at optical character recognition and document understanding in Indian scripts, used with a layout parser and a reading-order network.^[53] Version 2.1 added key-value extraction from forms, Indic handwriting recognition and better handling of complex tables.^[43] Sarvam Edge, announced in February 2026, packages compact speech recognition, speech synthesis and translation models to run on devices without a network connection.^[56]

Products and deployments

Sarvam sells its models through APIs and through applications built on them, and offers deployment on its own cloud, in a customer's private cloud, or on-premises in air-gapped environments.^[1] At Epoch it described Indus as a single workspace for several agent products: Voice Agents for calls, Work Agents for employee tasks, Document Agents, Content Studio for dubbing and voice content, and coding agents.^[40] Earlier products include Samvaad for conversational agents, Arya for orchestrating agents, and Akshar, a document intelligence workbench built on Sarvam Vision.^[10][31]

In the June 2026 funding announcement, Sarvam said its conversational platform handled more than 2 million interactions a day, its inference platform processed 10 million API calls a day, its speech models transcribed more than half a million hours of audio a month, and Sarvam Vision was being used to digitise more than 35 million pages.^[4] It also said its voice agents had collected data from 17 million farmers for the Ministry of Agriculture and Farmers' Welfare and had supported renewals for 45 million policyholders of a large insurer.^[4] Named customers include UIDAI and Tata Capital; NVIDIA's case study also names Infosys.^[23][1][26]

Funding

Separately, the government's compute support was structured as compulsorily convertible debentures. Inc42, citing The Economic Times, reported in June 2026 that these were expected to convert into a government stake of 1% to 2% once the Series B closed, and quoted an unnamed official saying that support under the IndiaAI Mission "needs to be accounted for in some form, if not cash".^[8] The same report said several startups had pushed back against the government taking equity and favoured grants instead.^[8]

As of October 2026, Sarvam's website listed Lightspeed, Peak XV Partners, Khosla Ventures, HCLTech, Bessemer Venture Partners and NVIDIA as investors.^[2] Amazon and Prosperity7 Ventures, named in Bloomberg's April report, did not appear in the June announcement or on that list.^[39][4][2] Inc42's company-data page lists Sarvam's total revenue for the year to March 2025 as 29.1 crore rupees, up from 11.1 crore rupees the year before, with a net loss of 102.4 crore rupees.^[57] Separately, Inc42 reported in June 2026 that, according to unaudited financials shared by HCLTech, Sarvam's revenue was 1.5 crore rupees in the year to March 2025 and 45.1 crore rupees in the year to March 2026.^[8]

Reception and criticism

Sarvam-M drew a sharp reaction in May 2025. Deedy Das, an investor at Menlo Ventures, wrote on X that "India's biggest AI startup" had launched "a 24B Mistral small post trained on Indic data with a mere 23 downloads 2 days after launch", contrasted it with an open model from two Korean students that had about 200,000 downloads in a month, and called it "embarrassing".^[58] Zoho co-founder Sridhar Vembu defended the company, writing on X: "Instant success is neither necessary nor sufficient to succeed long term."^[58] Business Standard later wrote that Sarvam-M's reliance on a foreign base model had raised questions about whether it could be called a sovereign model, and presented the from-scratch 105B model as a response to that criticism.^[59]

After the 105B model was unveiled, critics on X examined its configuration file and called it a "scaled-down DeepSeek architecture clone"; Kumar posted on X that his team "admire the Deepseek team and follow and learn from their research".^[60] An opinion piece in Swarajya, which reported the exchange, argued that sharing an architecture is not copying a model, pointing to Sarvam's own Indic tokenizer, training data pipeline and reinforcement learning infrastructure.^[60]

Writing in Forbes, the analyst Janakiram MSV called the 105B model "a genuine engineering achievement" and "not a wrapper", but noted that every performance figure was self-reported: at the time of the weight release there was no arXiv paper, no listing on the Hugging Face Open LLM Leaderboard or LMSYS Arena, and the Indian-language benchmark had been designed by Sarvam and judged by Gemini on prompts Sarvam selected.^[37] He argued that India lacked an independent institution to verify sovereign models that were already being deployed in public services.^[37]

Consumer uptake of Indus was modest at first. TechCrunch, citing Appfigures, reported in May 2026 that the app had been downloaded just over 293,000 times in India in the nearly three months since launch, against 43.9 million downloads of ChatGPT in the country.^[36]

See also

• List of neolabs
• Neolab
• Sovereign AI
• India AI Impact Summit 2026
• Nemotron
• Voice AI

References

1. ^1 ^2 ^3 ^4 ^5Sarvam. "India's Full-Stack Sovereign AI Platform" (company homepage and site footer). Retrieved 27 October 2026.
2. ^1 ^2 ^3 ^4 ^5 ^6Sarvam. "About us". Sarvam AI. Retrieved 27 October 2026.
3. ^1 ^2 ^3 ^4 ^5 ^6 ^7Manish Singh. "Five-month-old Indian AI startup Sarvam scores $41M funding". TechCrunch, 7 December 2023.
4. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11HCLTech. "Sarvam raises $234 million in first close of $300 million Series B at $1.5 billion valuation". Press release, 15 June 2026.
5. ^1 ^2 ^3 ^4NVIDIA. "NVIDIA Launches Nemotron Coalition of Leading Global AI Labs to Advance Open Frontier Models". NVIDIA Newsroom, 16 March 2026.
6. ^1 ^2 ^3Press Information Bureau, Ministry of Electronics and IT. "India's Common Compute Capacity Crosses 34,000 GPUs: 3 More Startups to Develop and Deploy Indian Foundation Models". PIB Delhi, 30 May 2025.
7. ^1 ^2Sarvam. "The government of India selects Sarvam to build India's sovereign large language model". Sarvam blog, 26 April 2025.
8. ^1 ^2 ^3 ^4 ^5 ^6Inc42. "Centre May Acquire 1-2% Stake In Sarvam Post $300 Mn Raise: Report". Inc42, 25 June 2026.
9. ^1 ^2 ^3 ^4Jagmeet Singh. "Indian AI lab Sarvam's new models are a major bet on the viability of open source AI". TechCrunch, 18 February 2026.
10. ^1 ^2 ^3 ^4 ^5 ^6 ^7 ^8 ^9 ^10 ^11 ^12 ^13 ^14 ^15Sarvam. "Open-Sourcing Sarvam 30B and 105B". Sarvam blog, 6 March 2026.
11. ^1 ^2 ^3 ^4Jagmeet Singh. "Sarvam becomes India's newest AI unicorn with $234 million funding round led by HCLTech". TechCrunch, 15 June 2026.
12. ^1 ^2 ^3 ^4 ^5Entrackr. "Exclusive: Sarvam AI board to approve $74 Mn funding from NVIDIA, Glade Brook, others". Entrackr, 3 August 2026.
13. ^1 ^2 ^3ETtech (The Economic Times). "Homegrown artificial intelligence model startup #Sarvam is in talks to raise another $250-300 million at a $3 billion valuation...". Post on X, 13 October 2026.
14. ^1 ^2 ^3 ^4Inc42. "Sarvam To Build Trillion-Plus AI Model In India, Launches Inference Service". Inc42, 30 July 2026.
15. ^1 ^2 ^3 ^4Sarvam. "Sarvam AI secures a landmark $41 million in Series A funding". Sarvam blog, 7 December 2023.
16. ^1 ^2Analytics Vidhya. "Indian Startup Releases OpenHathi: First-ever Hindi LLM". Analytics Vidhya, 14 December 2023.
17. ^1 ^2Sarvam AI. "OpenHathi-7B-Hi-v0.1-Base" model card. Hugging Face, December 2023.
18. ^1 ^2 ^3Sarvam. "Sarvam product launch: Sarvam Launches Revolutionary Full-Stack Generative AI Platform for India". Sarvam blog, 14 August 2024.
19. ^1 ^2Sarvam AI. "shuka-1" (Shuka v1) model card. Hugging Face, August 2024.
20. ^Sarvam. "Giving voice to India's linguistic diversity with Bulbul". Sarvam blog, 23 August 2024.
21. ^1 ^2 ^3Sarvam. "Sarvam 1: The first Indian language LLM". Sarvam blog, 24 October 2024.
22. ^1 ^2 ^3Sarvam AI. "sarvam-1" model card. Hugging Face, October 2024.
23. ^1 ^2 ^3 ^4Press Information Bureau, Ministry of Electronics and IT. "UIDAI partners with indigenous GenAI Company Sarvam AI to enhance user experience of Aadhaar services". PIB Delhi (Release ID 2112485), 18 March 2025.
24. ^Sarvam. "Aadhaar activates voice agents for 4 crore Indians". Sarvam customer story. Retrieved 27 October 2026.
25. ^Sarvam. "Building a sovereign AI ecosystem for India". Sarvam blog, 20 May 2025.
26. ^1 ^2NVIDIA. "Sarvam AI Brings Sovereign AI to 1.4 Billion People With NVIDIA". NVIDIA customer case study. Retrieved 27 October 2026.
27. ^1 ^2 ^3Sarvam. "Sarvam-M: Explorations in Post Training and Inferencing Optimizations for a Hybrid Indic LLM". Sarvam blog, 23 May 2025.
28. ^1 ^2Sarvam AI. "sarvam-m" model card. Hugging Face, May 2025.
29. ^1 ^2Sarvam AI. "sarvam-translate" model card. Hugging Face, June 2025.
30. ^1 ^2Business Today Desk. "Meet the minds behind Sarvam AI: How Pratyush Kumar and Vivek Raghavan are building India's sovereign AI stack". Business Today, 10 February 2026.
31. ^1 ^2Sarvam. "Blog: Latest news" (index of dated posts) and RSS feed. Retrieved 27 October 2026.
32. ^1 ^2Sarvam. "Sarvam Announces Sovereign AI Partnerships with Indian States". Sarvam blog, 9 February 2026.
33. ^Ivan Mehta. "India's Sarvam wants to bring its AI models to feature phones, cars, and smart glasses". TechCrunch, 18 February 2026.
34. ^Jagmeet Singh. "India's Sarvam launches Indus AI chat app as competition heats up". TechCrunch, 20 February 2026.
35. ^Sarvam. "Introducing Indus". Sarvam blog, 20 February 2026.
36. ^1 ^2Ivan Mehta. "Finnish phone maker HMD bundles Indian AI chatbot onto new smartphone in push to reach local market". TechCrunch, 21 May 2026.
37. ^1 ^2 ^3 ^4Janakiram MSV. "India Can Train A Sovereign Model But Still Cannot Prove It Works". Forbes (contributor), 7 March 2026.
38. ^1 ^2Storyboard18. "Sarvam AI eyes $250 million raise at $1.5 billion valuation with Nvidia, Accel and HCLTech in talks". Storyboard18, 25 March 2026.
39. ^1 ^2 ^3Bloomberg. "India AI Startup Sarvam Raises Funds at $1.5 Billion Valuation". Bloomberg, 2 April 2026.
40. ^1 ^2 ^3 ^4Sarvam. "Everything we announced at Sarvam Epoch". Sarvam blog, 7 August 2026.
41. ^1 ^2Free Press Journal. "Sarvam AI Onboards Nvidia As Strategic Investor In $75 Million Series B Extension, First Indian Firm To Do So". Free Press Journal, 5 August 2026.
42. ^Sarvam. "Introducing Sarvam Circle". Sarvam blog, 5 August 2026.
43. ^1 ^2 ^3Sarvam. "Sarvam Vision 2.1: Pushing the Pareto frontier of document intelligence". Sarvam blog, 24 September 2026.
44. ^1 ^2 ^3Sarvam. "Introducing Saaras V4". Sarvam blog, 25 September 2026.
45. ^1 ^2 ^3Sarvam. "Introducing Bulbul V4 Flash". Sarvam blog, 13 October 2026.
46. ^CXO Digitalpulse. "Sarvam AI Funding: Startup Eyes $300 Million at $3 Billion Valuation". CXO Digitalpulse, 13 October 2026.
47. ^1 ^2 ^3Indian Institute of Technology Madras, Department of Computer Science and Engineering. "Pratyush Kumar" faculty page (archived 21 March 2023).
48. ^Storyboard18. "The minds behind Sarvam AI: IIT alumni building India's sovereign language model". Storyboard18, 2026.
49. ^Gowtham Ramesh, Sumanth Doddapaneni, et al. (including Vivek Raghavan, Anoop Kunchukuttan, Pratyush Kumar and Mitesh M. Khapra). "Samanantar: The Largest Publicly Available Parallel Corpora Collection for 11 Indic Languages". arXiv:2104.05596, April 2021.
50. ^Jay Gala, Pranjal A. Chitale, et al. (including Vivek Raghavan, Pratyush Kumar, Mitesh M. Khapra, Raj Dabre and Anoop Kunchukuttan). "IndicTrans2: Towards High-Quality and Accessible Machine Translation Models for all 22 Scheduled Indian Languages". arXiv:2305.16307, May 2023.
51. ^1 ^2Sarvam AI. "sarvam-30b" model card. Hugging Face, March 2026.
52. ^1 ^2Sarvam AI. "sarvam-105b" model card. Hugging Face, March 2026.
53. ^1 ^2Sarvam. "Sarvam Vision". Sarvam blog, 5 February 2026.
54. ^1 ^2Sarvam. "Introducing Saaras V3: Built for the Way India Speaks". Sarvam blog, 11 February 2026.
55. ^1 ^2Sarvam. "Introducing Bulbul V3: Natural. Expressive. Production-ready.". Sarvam blog, 7 February 2026.
56. ^Sarvam. "Announcing Sarvam Edge". Sarvam blog, 14 February 2026.
57. ^Inc42. "Sarvam AI: Financials, Revenue, P&L Statement". Inc42 company data. Retrieved 27 October 2026.
58. ^1 ^2Outlook Business. "Zoho's Sridhar Vembu Weighs In on Sarvam-M: 'Instant Success Is Neither Necessary'". Outlook Business, 26 May 2025.
59. ^Abhijeet Kumar. "Why Sarvam's new 105B model marks a shift in India's sovereign AI ambitions". Business Standard, 19 February 2026.
60. ^1 ^2Raghavan S Rao. "Why calling Sarvam a DeepSeek clone gets the story of AI exactly backwards". Swarajya, 7 March 2026.