Sarvam Raised $234 Million at a $1.5 Billion Value and Released Two Open Models, but the Government's August List of Launched Models Names Neither Soket AI nor Gan.ai, Which Share Rs 287 Crore of Approved Support

Sarvam AI announced on 15 June 2026 that it had closed $234 million of a planned $300 million Series B at a post-money valuation (the company’s worth just after the new money) of $1.5 billion. A Lok Sabha reply reported by ANI on 6 August 2026 names launched models from Sarvam and from Gnani.ai, and none from Soket AI or Gan.ai, the other two startups chosen in 2025 to build foundation models (the large AI models behind chatbots) under the IndiaAI Mission.

The mission’s budget, GPU count and full list of twelve funded builders are in tuput’s IndiaAI Mission piece, and the case for language-specific models is in India is building AI in its own languages. This one follows the companies themselves: who has raised what, what each has put in public, and who measured it.

The four chosen startups, side by side

PIB’s release of 30 May 2025 says Sarvam was selected on 26 April 2025 to build an open-source 120-billion-parameter model. Parameters are the adjustable numbers inside a model, a rough gauge of size. The same release announced three more startups, Soket AI, Gnani.ai and Gan.ai. The annexure the government tabled in the Rajya Sabha on 13 February 2026 lists the support each received.

The four add up to Rs 711.10 crore by tuput’s addition. The government says it covers the actual cost of the compute the teams use and sets aside a further 25 per cent for datasets and staff. The annexure shows no non-compute money for Sarvam or Gnani.ai.

Sarvam: $349 million in disclosed funding

Sarvam’s first money, $41 million across seed and Series A, was reported in December 2023. In June 2026 HCLTech put in $150 million as lead strategic investor, Bessemer Venture Partners joined, and Khosla Ventures and Peak XV Partners stayed in. Sarvam says the funds go to research on its next model, aimed at agentic, coding and cybersecurity uses, plus compute and work placed inside client companies.

Entrackr reported on 3 August 2026, from a filing with the Registrar of Companies, an extension of about Rs 698 crore ($74 million) led by Nvidia at Rs 238 crore ($25 million), with Glade Brook Capital at Rs 190.3 crore ($20 million). Entrackr puts the value at $1.51 billion and says Nvidia would hold 1.66 per cent and HCLTech 9.95 per cent. tuput found no Sarvam statement on the extension. The 2023 money, the June close and the reported extension add to $349 million.

Sarvam’s own usage figures are 10 million API calls a day, speech models transcribing over 500,000 hours of audio a month and over 2 million interactions a day on its conversational platform. TechCrunch repeated them and noted that no revenue figures were disclosed.

What Sarvam released, and who scored it

Sarvam’s post of 6 March 2026 announced a 30-billion and a 105-billion-parameter model. Both are mixture-of-experts designs, meaning only part of the network runs for each word. The 30B uses 2.4 billion active parameters, and the 105B’s model card says 10.3 billion. The post says both were trained in India on IndiaAI compute, on 16 trillion and 12 trillion tokens, and that the weights are on AIKosh, the government’s model and dataset platform, and on Hugging Face under the Apache 2.0 licence. Open weights means anyone can download the trained numbers and run them. The model Sarvam was selected for had 120 billion parameters, and the one it released has 105 billion.

On the model card, Sarvam compares the 105B with GLM-4.5-Air, gpt-oss-120B and Qwen3-Next-80B. It leads on BrowseComp, a web-research test, at 49.5 against 21.3 and 38.0 for two of them. It trails on SWE-Bench Verified, a software-fixing test, at 45.0 against 57.6, 50.6 and 60.9. The card does not say who ran the tests. The blog says Gemini 3 judged the Indian-language benchmark.

An outside scorer, Artificial Analysis, gave the 105B an Intelligence Index of 18 on 2 April 2026, behind gpt-oss-120B at 33, and said both models guess more than they decline. That result is covered in the IndiaAI piece, and Sarvam’s speech scores are in the voice and translation roundup.

Gnani.ai: a 5-billion-parameter preview and no published score

PIB’s May 2025 release said Gnani.ai would build a 14-billion-parameter voice model. At the India AI Impact Summit in February 2026 it showed Inya VoiceOS, a 5-billion-parameter model that takes speech in and gives speech out without a text step in between. Open Source For You, reporting on 19 February, calls it a research preview that comes before the planned 14B model. The article gives over 14 million hours of speech in pretraining, more than 15 Indian languages, sub-second response time and 24 kHz audio output. It gives no benchmark, no licence and no availability, and Gnani.ai’s chief executive is quoted claiming better accuracy than older speech-to-text-then-text-to-speech chains, with no figure. tuput found no report of the 14B model’s release.

Gnani.ai’s Vachana speech-to-text model, trained on over 1 million hours of voice data according to Inc42, is a separate product. On money, Inc42 reported on 31 March 2026 a $10 million Series B (about Rs 94 crore) led by Aavishkaar Capital, with Info Edge Ventures also in, and the company claims more than 200 enterprise customers.

Krutrim: outside the programme, and out of the model race

PIB’s list of twelve does not include Krutrim, the AI company founded by Ola’s Bhavish Aggarwal, or CoRover. Krutrim raised $50 million at a $1 billion valuation in January 2024. MediaNama reported on 6 February 2025 that it had launched Krutrim-2, a 12-billion-parameter model, and Ola announced a Rs 2,000 crore (about $230 million) investment in an AI lab.

The model card says Krutrim-2 is built on Mistral’s NeMo architecture, with a 128,000-token context window, under the Krutrim Community License rather than Apache. Krutrim ran its own tests with the LM Evaluation Harness. It scored 0.63 on MMLU, a general-knowledge exam, against 0.68 for Mistral’s NeMo Instruct, and 0.95 on IndicSentiment against 0.70.

On 5 May 2026 TechCrunch reported that Krutrim was moving from model development to cloud services. The Kruti assistant app had been pulled from app stores in April, chip design was paused in a late-2025 overhaul, and local media had reported more than 200 roles cut. Krutrim reported FY2026 revenue of about Rs 300 crore and its first annual net profit. About 90 per cent of FY2025 revenue had come from group companies, according to earlier reports, and Krutrim did not disclose the FY2026 split. The analyst Sanchit Vir Gogia of Greyhound Research told TechCrunch that “The standard of proof must rise with the claim.”

CoRover: a score its authors ran themselves

CoRover.ai, which makes BharatGPT, is not on PIB’s list either. Google said on 16 December 2025 that it was giving startups including CoRover and Gnani.ai $50,000 each and that they were using its Gemma open models for voice and e-governance models.

On 24 April 2026 CoRover published results for BharatGPT mini, a 0.5-billion-parameter model that runs on a device without a GPU. It reports an efficiency index of 64.6 for the mini and 1.7 for Sarvam 30B. The page defines the index as the average benchmark score divided by parameter count, so a small model gets a high number by construction. The post calls its benchmark independent, while its disclaimer says the scores are the authors’ own evaluation runs.

Gan.ai and Soket AI: approved money, models still to come

Gan.ai was founded in 2021 and sells personalised video. TechStartups reported a $5.25 million seed round in May 2023, led by Surge and Sequoia Capital India. The mission support is Rs 110.03 crore, of which Rs 88.02 crore is compute, for a 70-billion-parameter model aimed at text-to-speech. Gan.ai is not among the launches in the Lok Sabha reply, and tuput found no published model or later funding report.

Soket AI has Rs 177.08 crore in support, of which Rs 162.47 crore is compute. Inc42 reported on 22 July 2026 that a 24-billion-parameter model meant to prove the architecture is expected by the end of 2026 and will go first to a small group of enterprise partners. The 120-billion-parameter model would follow in 2027, and both are meant to be open-sourced. Soket has released an Indic pretraining dataset, has not disclosed revenue, and is looking to raise money in amounts it has not disclosed. Its founder, Abhishek Upperwal, told Inc42, “We want to do some amazing work on the research front first.” Inc42 says commercial deployments should begin gradually in 2027.