Sarvam’s AI Arsenal

In June, Sarvam AI entered India’s unicorn club, raising $234 Mn (about ₹2,210 Cr) in a $300 Mn Series B round at a $1.5 Bn valuation. Just a month later, the startup made it clear that the funding was only the beginning of a much bigger ambition.

At its inaugural developer conference, Sarvam Epoch, the startup unveiled its vision that it was no longer content with being known as a builder of Indian-language foundation models. Instead, it now wants to be a full-stack AI venture, providing everything India Inc needs to build, deploy and scale AI — from foundation models and infrastructure to coding copilots, voice agents, document intelligence and workplace applications.

This is a major shift from OpenHathi-v1, the open-source Hindi language model. Today, the startup is building across every layer of the AI stack, spanning foundation models, infrastructure and applications for speech, vision, coding, cybersecurity and defence.

What’s more interesting is that its ambitions just don’t stop at software. The Indian AI giant now has hardware, such as smart glasses (Sarvam Kaze), on the cards.

But the real question is why it is betting on the entire AI stack. Besides, can it become the new operating system for Indian enterprises that are already dependent on global players like OpenAI and Anthropic? Let’s try to answer these questions in this edition of The Outline.

Sarvam’s Sales Pitch

At its core, Sarvam knows that just building language models for India won’t cut it, and it has to make a bigger bet. To realise this, it has launched Sarvam Inference, an India-hosted platform for running frontier open-weight models, for enterprises that want to deploy AI while keeping sensitive data within India.

The platform currently supports Sarvam’s own 105 Bn-parameter model as well as models such as GLM 5.2 and Gemma 4. Sarvam is also building a trillion-parameter frontier model in India. 

However, Sarvam’s larger ambition is to capture a significant share of India’s AI consumption. To pursue this and boost its frontier AI push, the Indian AI startup unicorn has appointed Devendra Singh Chaplot, a founding member at Mistral AI and Thinking Machines Lab, and former xAI pre-training lead. It has also opened a new San Francisco office as part of its global expansion.

Besides, the startup is betting on competitive pricing, reliability and data security and sovereignty to lock horns with its global counterparts. But Sarvam wants to do something that sounds simple on paper and is extremely difficult in practice: convince Indian businesses to switch from OpenAI, Anthropic or Google for their AI stack.

Sarvam’s Show Of Strength: Will Indian Enterprises Buy In?

At its event, Sarvam showed on various fronts that its technology is competitive. For starters, its coding agent, Sarvam Code, performed strongly on several benchmarks. Then, its speech model, Saaras V4, supports all 22 constitutional Indian languages. Its voice platform, Samvaad, has already handled 325 Mn minutes of customer conversations over the past year.

Not just this, its work agents can process large amounts of company data, while its inference platform can run open models from infrastructure located in India. Upping the ante, Sarvam is also building a trillion-parameter model and is scaling its computing capacity to 10,000 accelerators.

Sarvam’s latest event was also packed with benchmark results:

• Sarvam Code scored 72 out of 89 on Terminal Bench, ahead of Claude Code using several models and only slightly behind Codex in Sarvam’s comparison. 
• On Data Agent Bench, its score rose from 61.4 with the standard model setup to 82.1 after Sarvam added its own skills and workflow. It also reported completing all 41 tasks on Exploit Bench.
• Saaras V4 claims the best average performance across seven standard English benchmarks while also supporting India’s 22 constitutional languages. Bulbul V4 brings more control over voice expression.
• Sarvam Work scored 79% on HarnessBench.

But enterprise buying decisions are rarely made on benchmark scores alone. Benchmarks tell you how a system performed on a particular test. They do not tell you whether a large enterprises will trust it with its most important workflows.

This matters particularly for products such as coding agents. An engineering team may not choose between two numbers on a leaderboard. It has developers who have already built habits around tools such as Claude Code and Codex, which have become part of the way teams write, test and review software.

One AI founder put the problem bluntly: Codex and Claude Code are already “absorbed into the workflow. For an established team, moving to something new is not necessarily worth the disruption”.

He added that a developer may be impressed by Sarvam Code due to its lower cost, but a CIO is likely to look beyond the benchmark scores and ask a more basic question: does the product fit into its existing systems and workflows, and offer enough value to justify moving away from tools that employees already use?

Sarvam cannot simply tell enterprises that its models are cheaper or match competitors on benchmark scores. It needs to offer a far more compelling reason for businesses to switch. And that is where Sarvam’s real test begins. 

The Ace Up Sarvam’s Sleeve

Sarvam has advantages that no one can deny. For one, Sarvam Inference allows developers and businesses to run open models from infrastructure hosted in India. It is building its computing capacity within the country and says it wants to produce the largest share of the tokens India consumes right here. This matters for government departments, defence organisations or highly regulated businesses.

Sarvam’s Anvaya platform, for instance, is being built for defence, national security and intelligence use cases and can run entirely on-premise. Then, for some organisations, keeping sensitive information within India might be a requirement. And that’s precisely where Sarvam has an edge. 

Sarvam is also building products around India’s particular linguistic complexity. Saaras V4 supports all 22 constitutional languages, including languages that have very little data available for training. Samvaad has already handled hundreds of millions of minutes of real customer conversations, including conversations where people switch between languages.

On the other hand, Google, Meta and other global companies are still investing in this area. Taking on Sarvam in its home arena could prove difficult for its peers until they offer on-par Indian-language capabilities at competitive pricing.

Sarvam’s pitch on pricing is strong. Its Samvaad voice platform costs ₹3.5 a minute, compared with an industry range of roughly ₹8-12, according to the startup. Its 105B model is positioned as significantly cheaper than comparable models from Google and OpenAI. Sarvam Code’s cost-per-task comparison is even more striking.

Inside Sarvam’s Full-Stack Bet

Sarvam’s answer to the global AI giants appears to be its full-stack approach. It does not just want to sell a model but wants to control several layers around it.

Sarvam Inference sits at the infrastructure layer, then comes the model layer with the likes of Sarvam 105B, Sarvam Vision 2.0, Saaras and Bulbul. To wrap it all up, Sarvam’s Indus platform offers agents for work, voice, content, documentation and coding.

In other words, Sarvam Code is for developers; Sarvam Work for workplace tasks; Samvaad for voice agents; Vision for documents and Anvaya for sensitive intelligence use cases. They also have a new Kivi desktop app for dictation. This is an alternative to the likes of Wispr Flow.

The idea is that these products can work together, giving Sarvam an advantage that is difficult to see in a benchmark. If it controls the model, the infrastructure and the software around it, it can optimise the entire system for cost and performance.

Sarvam is betting that the future enterprise AI buyer will not necessarily want the world’s best model for every task. It may want the right model for each task, at the right price, with the right level of security. And that’s the market Sarvam is looking at. 

[Edited by Shishir Parasher]

The post Sarvam’s AI Arsenal appeared first on Inc42 Media.