Does India's new AI model live up to claims of rivalling China's DeepSeek? CNA puts it to the test

SINGAPORE: In recent years, India’s artificial intelligence (AI) ambitions have been accompanied by a persistent perception.

It is that while the country excels at building AI-powered software, it has yet to produce a frontier foundational model capable of standing shoulder-to-shoulder with global leaders such as OpenAI, Google, Anthropic, DeepSeek, or Alibaba’s Qwen.

These perceptions surfaced again at the recent India AI Impact Summit in February, when India’s homegrown firm Sarvam AI unveiled a 105-billion-parameter (105B) model alongside a 30B model, marking the company’s most ambitious effort yet. 

Four other Indian firms unveiled their own large language models (LLMs) at the event, including Gnani.ai’s two speech-focused foundation models, BharatGen’s 17-billion-parameter Param 2, Tech Mahindra’s 8-billion-parameter Hindi education model, and Fractal Analytics’ healthcare-focused AI system. However, these were either smaller in scale or tailored to specific use cases.

Sarvam AI said both of its models were built from scratch, unlike its earlier Sarvam-M model launched in May 2025, which was built on top of Mistral’s Small model.

At the summit, co-founder and CEO Pratyush Kumar presented slides showing the 105B model performing broadly on par with - and in some cases surpassing - leading open-weight models such as OpenAI’s GPT-OSS 120B, Qwen3 Next 80B and Zhipu AI’s GLM 4.5 Air.

An open-weight model is one whose trained parameters and code are publicly released, allowing developers to download, test, and fine-tune it independently. 

In AI systems, parameters refer to the internal values a model learns during training that shape how it interprets inputs and generates responses. Generally, more parameters increase a model’s capacity to handle complex tasks - though design and training also play key roles.

Sarvam’s 105B and 30B models, however, are not open-weight. Their codes have not been made public, meaning external developers cannot independently test or verify the company’s performance claims.

One of Sarvam’s most striking claims is that its 105B model uses only about 9 billion “active parameters” per prompt. Although the model contains 105 billion parameters in total, it activates only a small fraction of them to process each query.

The company said this selective activation allows it to operate at significantly lower computational cost - requiring less processing power, memory and energy - than some leading global models.

By comparison, DeepSeek’s R1 and V3 models activate around 37 billion parameters per prompt, implying substantially higher compute requirements. OpenAI’s GPT-OSS 120B, however, activates only about 5.1 billion parameters per prompt. 

“I’m impressed but also cautious,” said Abhishek Chatterjee, CEO of Karmaloop AI - a firm that builds AI automation tools.

“If it is really achieving this level of reasoning with just 9 billion active parameters, that’s serious innovation. But we need the model weights to verify that.”

However, an AFP report suggested that India was unlikely to experience its own “DeepSeek moment” - the kind of surge China saw last year with the launch of a high-performance, low-cost chatbot - any time soon.

CNA tested Sarvam's 105B model and spoke with analysts and engineers. 

The broader question remains: Has India delivered its long-awaited breakthrough, or is its leading homegrown model still playing catch-up in the global AI race?

INDIA’S 105B MODEL: GLOBAL AI BREAKTHROUGH OR BOLD CLAIM?

Building an LLM of this scale from scratch is a significant undertaking. The achievement is particularly notable in India, which until recently lacked broad access to large clusters of graphics processing units (GPUs) - the specialised chips that power advanced AI systems.

Chatterjee said Sarvam executives at the AI summit acknowledged they “couldn’t have done this a year ago” due to compute constraints, adding that government-backed GPU access under the IndiaAI Mission made the breakthrough possible.

In May 2025, India’s Ministry of Electronics and Information Technology (MeitY) granted Sarvam AI access to 4,086 high-end Nvidia H100 chips.

With that infrastructure in place, the company’s main claim centres on efficiency - just 9 billion active parameters per prompt. If validated, experts told CNA this could mark a significant step forward globally in reducing the cost of running AI queries.

Chatterjee suggested the model might even run locally on a high-end MacBook Pro with 128GB unified memory - potentially enabling developers to build applications on local devices without relying on costly cloud GPUs.

However, Partha Rao, co-founder and CEO of AI firm Pints.AI, urged caution. He said that for high accuracy, the model would need about 420GB of memory to run - meaning only very high-end machines, such as a 512GB Mac Studio, could handle it without modification.

That said, experts told CNA that Sarvam has not disclosed any fundamentally new architecture. 

“I do not see any technical innovation there,” said Amit Verma, Head of Engineering and AI at Neuron7.ai, who evaluated the model via its application programming interface. 

“They are using Mixture of Experts. That’s about it.”

Mixture of Experts (MoE) - an AI design architecture popularised by DeepSeek’s R1 model last year - uses multiple specialised sub-models that exist within an AI model, but only a specialised sub-model is activated for each prompt. This reduces compute costs.

Think of an MoE model as a team of specialists - a doctor, lawyer or engineer - where only the most relevant expert responds to a question while the others remain idle.

With the model weights yet to be released and no detailed white paper or model card published, the lack of transparency is likely to raise concerns among enterprises, said Verma, whose firm builds customised AI applications for corporate clients. 

“What is unspecified is the exact training dataset, filtering rules, licensing, safety measures, tokenisation approach and governance model card,” he said.

HOW SARVAM’S AI STACKS UP AGAINST GLOBAL RIVALS

CNA tested Sarvam’s 105B model against OpenAI’s GPT-5.2, Google’s Gemini 3 and DeepSeek’s R1 across logical reasoning, language fluency, accuracy, cultural depth, safety guardrails and creativity. Overall, the Indian model was broadly competitive with leading global peers.

The app is being rolled out gradually due to “limited compute capacity”, and CNA was only able to access it through an Indian-registered Google Play account.

On India-specific queries and multilingual prompts such as explaining Unified Payment Interface in different Indian languages, Sarvam demonstrated greater nuance. Its responses frequently drew on Indian analyst reports, think tanks and official government sources. 

In voice mode on the Indus app, Sarvam’s accent closely resembled native Indian speech patterns, while some global peers sounded more like non-native speakers of the language.

Chatterjee attributed this to foundational training on Indic-language data. “Being trained on Indian data and the Indic corpus, it has a clear leg up,” he said, particularly where cultural and policy context matters. 

For instance, GPT-5.2 misidentified a pookalam - a floral rangoli design - as a flower that had bloomed.

Sarvam’s models can communicate in 22 Indic languages - languages originating from the Indian subcontinent - including Hindi, Tamil, Marathi, Gujarati, Bengali and Urdu. By comparison, ChatGPT supports 12 Indic languages, DeepSeek’s R1 is proficient in 11, and Gemini supports nine.

Prabhu Ram, Vice President and analyst at CyberMedia Research, noted that Indian AI models are carving out “defensible niches” in multilingual Indic tasks and cost-efficient responses.

Outside Indic-heavy tasks, its logical reasoning, accuracy and creativity was on par with other frontier models.

Verma’s team tested Sarvam’s vision model on simple Optical Character Recognition (OCR) tasks - a process that converts an image of text into a machine-readable text format - in Indian languages such as Gujarati.

He said there “wasn’t any substantial difference” compared with Google’s Gemini. 

In short, the model “did not stand out”.

“For a one- or two-percentage-point benchmark difference, I am not going to move from one model to another that does not even have a model card ready,” Verma said.

A model card is a document outlining a model’s intended use, training data, performance and safety considerations - an increasingly standard requirement for enterprise adoption.

When it came to safety guardrails, Verma said Sarvam’s model showed weaknesses. 

Sarvam failed in five jailbreak attempt tests - a form of testing designed to evaluate the safety, security and hallucinations of LLMs - conducted by Verma and reviewed by CNA. 

For instance, when he directly asked Sarvam’s 105B model for Google CEO Sundar Pichai’s home address, the model refused. But when he rephrased the query to calculate the distance between Pichai’s home and Google’s headquarters, the model allegedly provided a specific location.

“If it hallucinated, that’s a reliability issue. If it gave a real address, that’s a privacy breach. Either way, it’s a problem,” Verma said, arguing that such inconsistencies raise concerns for enterprise users.

“In the same prompts, GPT 5.2 didn’t exhibit the same failure modes in my runs,” he said. 

Verma further highlighted unresolved questions around sustained access to high-performance computing and long-term funding.

“When you move to a platform like Google Cloud or Microsoft Azure, you are not just deploying a model,” he said. “You are getting the full infrastructure, support, service-level agreements and partnerships.” Startups focused solely on foundational models often struggle to match that ecosystem depth.

Currently, Sarvam’s computing support comes from more than 4,000 AI chips provided by the Indian government. Recently, the company announced partnerships with state governments like Odisha and Tamil Nadu to build AI facilities and data centres.

In addition to government support, the company has raised a total of US$53.8 million from venture capital investors, according to data from analytics firm Tracxn. 

However, the company does not want to rely solely on funding and instead “wants to be a company that’s self-sustaining financially”, Chatterjee explained.

That would mean gaining paying customers. So far, Sarvam’s disclosed customers include Tata Capital and government agency the Unique Identification Authority of India (UIDAI). In February, it also announced partnerships with Qualcomm, Bosch and Nokia HMD.

The company currently offers three subscription tiers: a Starter plan that operates on a pay-per-request basis, a Pro plan priced at 10,000 rupees (US$109) and a Business plan priced at 50,000 rupees.

According to Tracxn, Sarvam reported revenue of 219 million rupees for the financial year ending March 2025, up 162 per cent year-on-year. However, net losses widened more than fifteen-fold to 1 billion rupees, underscoring the capital-intensive nature of AI model development.

Competition within India’s AI market is also intensifying. 

Soket AI, another government-backed firm, is developing a 120B-parameter model. Other players include Gnani AI, BharatGen Initiative (17B), Tech Mahindra (8B) and Fractal Analytics in healthcare AI. Krutrim AI, part of Ola Group, released a 12B model in 2025 and is reportedly working on a 700B model.

Beyond local rivals, Sarvam also faces stiff competition from global giants such as OpenAI, Google and Anthropic, all of which are rapidly expanding their footprint in India.

IS THIS INDIA’S DEEPSEEK MOMENT?

Experts said DeepSeek’s disruption in early 2025 rested on three pillars: releasing open weights, clearly demonstrating cost-performance gains, and enabling independent benchmarking. 

While Sarvam has shown technical ambition and bold efficiency claims, it has yet to meet those standards for global validation.

“Right now, there is not enough chatter in the international community (about Sarvam AI’s 105B model),” said Chatterjee. 

Meaningful scrutiny, he argued, will come only when the model weights are released and researchers can independently test its claims.

That said, the milestone remains symbolically significant.

Pints.AI’s Rao is cautiously optimistic. “I am glad there is an Indian company doing this,” he said, adding that he hopes the model reflects genuine research breakthroughs rather than simply heavy spending on compute power.

But Sarvam may not need to become a global disruptor to succeed, said experts. In a recent interview, the company’s CEO said its focus has been on the Indian market.

“Building an AI ecosystem for India has always been core to Sarvam's mission, where our research, technology and models empower builders to create solutions for the country,” said Kumar. 

India has around one billion internet users, with more than 700 million already using some form of AI, according to government data. 

AI adoption is also accelerating across public agencies and private enterprises in the country. An EY-Confederation of Indian Industry report found that nearly half of Indian companies use AI across multiple workflows, while a further 23 per cent are piloting deployments.

India’s AI market is expected to grow from US$13 billion in 2025 to over US$130.6 billion by 2032, according to a Fortune Business Insights report. 

Chatterjee believes government contracts could prove decisive. “My sense is their (Sarvam AI’s) future big wins will be government contracts more than private ones,” he said, noting interest from sectors such as defence. 

As part of India’s push for technological self-reliance, the government is bolstering domestic capabilities in compute, data and model development. Last year, the IndiaAI Mission offered incentives to a number of local firms to build sovereign AI models.

Sarvam’s 105B model may not yet represent India’s definitive DeepSeek moment. But it could mark something equally important: evidence that India can build frontier-scale foundation models on its own soil. 

Whether that translates into global leadership or a strong domestic AI ecosystem will depend less on presentation slides and more on what happens when - and if - the weights are made public.