Sarvam AI launches 30B and 105B models, says 105B outperforms DeepSeek R1 and Gemini Flash on key benchmarks

• Sarvam AI unveils 30B and 105B parameter language models in India
• Both models use mixture-of-experts for higher efficiency
• 105B model outperforms larger global models on benchmarks

Did our AI summary help?

Bengaluru-based AI startup Sarvam AI on February 18 announced the launch of two new large language models, a 30-billion-parameter model and a 105-billion-parameter model, both trained from scratch, positioning them within India’s broader push to build sovereign AI capabilities.

The startup announced the models at the India AI Impact Summit in Delhi and said both models use a mixture-of-experts, MoE, architecture designed to improve efficiency while maintaining performance across reasoning, programming and tool-use tasks.

“Large language models, we have, of course, been building models across a 3-billion-parameter dense model, but it is important to scale up, and there are two other models that we trained and are talking about releasing today. One is Sarvam 30 Billion,” said Pratyush Kumar, cofounder of Sarvam.

The 30B model activates only 1 billion parameters per token, despite having 30 billion parameters overall.

“It is actually a mixture-of-experts model, we have a 30-billion-parameter model, but in generating every output token, it only activates 1 billion parameters,” Kumar said.

He said this structure reduces inference costs and improves efficiency in reasoning workloads.

“If you look at thinking budget, Sarvam 30B significantly outperforms both at the 8K and 16K scales compared to the latest models released at the same size.”

The model supports a 32,000-token context window and was trained on 16 trillion tokens. Kumar said efficiency remains central to the company’s thesis.

“We want to make AI work at population scale. Being able to do it efficiently becomes a very core thesis.”

105B model compared with global frontier systems

Sarvam also unveiled a 105-billion-parameter MoE model with 9 billion activated parameters and a 128,000-token context window, designed for more complex reasoning and agentic tasks.

“We trained a 105-billion-parameter model, it is also designed to do complex reasoning tasks very well,” Kumar said.

He compared the model’s performance with larger global systems.

“At 105 billion parameters, on most benchmarks this model beats DeepSeek R1 released a year ago, which was a 600-billion-parameter model.”

DeepSeek R1 is a reasoning-focused model released by Chinese AI firm DeepSeek and was reported to have 600 billion parameters, positioning it among the largest open models at the time of its launch. Gemini Flash, developed by Google, is a lighter and faster variant in the Gemini family, designed for lower latency and cost-efficient deployment compared to its larger frontier counterparts.

“It is cheaper than something like a Gemini Flash, but outperforms it in many benchmarks,” Kumar said.

On Indian language benchmarks, Kumar said the model delivers stronger performance than several larger competitors.

“Even with something like Gemini 2.5 Flash, which is a bigger and more expensive model, we find that the Indian language performance of this model is even better.”

The launch comes as India steps up efforts to build domestic large language models tailored to multilingual and large-scale public use cases, reducing reliance on foreign frontier systems.

Peak XV and Lightspeed-backed Sarvam AI  has emerged as a key player in India’s sovereign artificial intelligence (AI) push.

The IndiaAI Mission, backed by a Rs 10,000 crore fund to build India’s sovereign foundational model, has disbursed Rs 111 crore in GPU subsidies so far.

The biggest winner to date is Bengaluru-based Sarvam AI, which bagged a record 4,096 NVIDIA H100 SXM GPUs via Yotta Data Services, receiving nearly Rs 99 crore in subsidies.

Sarvam was earlier announced as the first startup selected to build India’s foundational AI model under the mission.

Founded in July 2023 by Vivek Raghavan and Pratyush Kumar, who previously worked at Infosys co-founder Nandan Nilekani-backed AI4Bharat, Sarvam develops a full-stack offering for Generative AI, ranging from research-led innovations in training custom AI models to an enterprise-grade platform for authoring and deployment.