Sarvam presents two new big language models with an emphasis on advanced reasoning and real-time usage

Sarvam Launches Two New Large Language Models

Bengaluru-based AI startup Sarvam on Wednesday launched two new large language models (LLMs) — Sarvam-30B and Sarvam-105B — as it expands into advanced reasoning and enterprise AI deployments.

The smaller model, Sarvam-30B, is built for efficient, real-time applications. It supports a context window of up to 32,000 tokens and has been trained on 16 trillion tokens. According to the company, the model is optimised for “efficient thinking,” meaning it can deliver stronger responses while using fewer tokens — helping reduce inference costs in production environments.

At launch, Sarvam shared benchmark comparisons of Sarvam-30B against models such as Gemma 27B, Mistral-32-24B, Qwen-30B and GPT-OSS-20B, among others. The evaluations covered tasks like Math500, HumanEval, MBPP, Live Code Bench v6 and MMLU, which measure mathematical reasoning and coding accuracy. The company said Sarvam-30B delivered competitive results across general reasoning and programming benchmarks.

On the AIME benchmark — which tests mathematical reasoning under different compute “thinking budgets” — Sarvam-30B showed improved performance as more compute was allocated, placing it alongside other 30B-class reasoning models.

Sarvam also introduced Sarvam-105B, a larger and more powerful model aimed at handling complex reasoning tasks. It supports a context length of up to 128,000 tokens. The company claims the model performs on par with several leading open and closed-source frontier models in its category.

The launch reflects Sarvam’s push to build foundational AI capabilities in India, at a time when startups are looking to reduce reliance on global APIs. As enterprises focus on cost efficiency, better control, and data residency, mid- to large-parameter open models are becoming a strong alternative for deployment.

Backed by Lightspeed and Peak XV Partners, Sarvam said the new models are designed for enterprise use cases such as coding assistance, research, analytics, and real-time AI agents. The company did not disclose pricing details.