Sarvam AI | All about the Bengaluru startup’s plans for a trillion-parameter model

Bengaluru-based startup Sarvam AI has announced plans to build a one-trillion parameter foundational model. To that end, the company has opened a San Francisco office and a Bay Area research lab, raising $234 million in the first close of its series-B funding round at a valuation of $1.5 billion.

Foundational models are general purpose systems capable of performing new tasks without explicit programming. However, their scope is predefined and determined by the datasets on which they are trained. These massive neural networks identify patterns within vast datasets to build a “foundation”, allowing them to apply learning to varied tasks such as coding, summarisation and image generation.

Foundational models are different from frontier models. The latter refers to systems from top-tier AI companies such as OpenAI, Anthropic, Moonshot AI and DeepSeek. These possess cutting-edge capabilities, including complex reasoning and high-end coding. All frontier models are foundational, but not vice versa.

As things stand, the scope of AI models remains within the territory of narrow or “weak” AI. But humans often overestimate these capabilities because of a natural tendency to seek patterns and anthropomorphise non-human entities.

Moreover, AI companies frequently hype their systems, claiming they are nearing artificial general intelligence (AGI) or strong AI — human-like versatility in reasoning and adapting across intellectual tasks.

AGI is the ‘X’ all the AI companies solve for. The north star of these companies is to turn intelligence into a utility, as Sam Altman said in so many words.

Where does Sarvam AI stand?

Currently, only a few countries have foundational models.

In 2017, Vladimir Putin said whoever controls AI will control the world. Today, AI dominates geopolitical discourse, with companies like Palantir and Elbit Systems deploying it in modern warfare. As the U.S. and China compete with bleeding-edge models, other nations are rushing to build infrastructure to stay in the fight.

The company recently launched a developer platform, Epoch Builder Edition, to help organisations build and deploy large language models (LLMs) tailored for Indian languages. The platform offers 7 billion and 70 billion parameter models trained on 2-trillion tokens and supports more than 10 Indian languages including Hindi, Tamil, Kannada and Bengali.

Sarvam has also partnered with IBM to develop sovereign AI technologies in India.

Challenges and advantages

Building a sovereign platform comes with significant challenges centered around hardware acquisition, elite engineering talent, and training datasets.

“All models are wrong, some are useful,” the British statistician George Box famously said in 1976.

To understand neural networks, we have to start with parameters.

When a model processes a prompt, it passes input through these parameters to predict the next word. The result is compared against the correct output to calculate error; through backpropagation and gradient descent, the network adjusts its internal settings to improve accuracy. The iterative process is known as pre-training.

Pre-training is expensive. Frontier models need tens of thousands of graphics processing units (GPUs) running in parallel for months, performing quadrillions of floating-point operations per second (FLOPS).

Single training runs can cost over $100 million in compute and energy. Silicon Valley giants and Chinese research institutions have hoarded the global chip stocks, creating a severe shortage for everyone else.

While giants like Microsoft and Google run clusters of over 1,00,000 GPUs, Sarvam AI operates on a much smaller capital footprint, making large-scale training runs highly risky.

Following pre-training, developers use supervised fine-tuning and reinforcement learning with human feedback (RLHF) to align model outputs with human preferences. These are then measured against global benchmarks.

Token tax

However, Sarvam AI possesses a major advantage in data efficiency.

Models don’t read human words. A tokenizer has to cut it up into small chunks or tokens. Western models are optimised for English text (one token is roughly 0.75 words). When these models process non-Latin scripts like Hindi, Bengali, or Malayalam, their tokenizers will need 8 tokens per word — a phenomenon known as ‘high token fertility.’

Sarvam has solved this bottleneck by building a custom tokenizer with a tailored vocabulary mapping common Indic subwords and phrases, reducing token fertility to just 1.2 to 2 tokens per word, leading to huge compute and inference cost savings.

However, Indian regional languages lack high-quality digitised text. While synthetic data may help, avoiding repetitive output is a challenge.

Foundational models are trained on trillions of tokens from web archives and literature. But raw internet data is often riddled with errors and duplication, meaning the cleaning process is prohibitively expensive.

Talent acquisition

Sarvam AI has recently hired Devendra Singh Chaplot as an advisor. He has worked with the world’s biggest AI labs including Facebook AI Research, Mistral AI and Thinking Machines Lab, and Elon Musk’s xAI. At the recent Epoch Builder Edition launch, he has rightly said how India’s unique languages and cultures warrants customised models built locally, rather than relying on imported Western technology.

Acquiring the necessary specialised researchers and engineers remains difficult as they are often recruited by Silicon Valley. Sarvam’s lower valuation compared to US and Chinese counterparts is a significant handicap in the global race for talent.

Efficiency gains

Chinese labs like Moonshot AI with its Kimi series, Alibaba with Qwen, and DeepSeek are releasing open weights. In other words, they are making the final, trained mathematical parameters of their models available for free download.

Thanks to open-weights, companies like Sarvam don’t have to spend millions of dollars to teach a network basic logic, world history and grammar from scratch. Instead, engineers can use knowledge distillation — a process where a smaller student model learns from a highly intelligent teacher model at a fraction of cost.

However, student models risk inheriting the fundamental biases and structural assumptions of their teachers.

To save on computing costs, Sarvam also uses a Mixture-of-Experts model; the AI doesn’t engage its entire brain for every single question. It uses a computational ‘gatekeeper’ to route the user’s question to only the specific ‘expert’ sub-networks needed for that exact task, making the system much faster and cheaper to run. Sarvam also uses grouped-query attention to minimise memory bandwidth overhead that lets a single graphics card serve multiple users at once.

Under strict frameworks like India’s Digital Personal Data Protection legislation, handling sensitive citizen records like health data, legal documents, or financial histories requires completely on-premise deployments. The data should also stay within local jurisdictions. For global, cloud-based public software providers will find it hard to accommodate such criteria. Sarvam AI can really cash in on the opportunity, and build for India.