Tech Explained: Here’s a simplified explanation of the latest technology update around Tech Explained: Sarvam 30B and 105B AI models are now open-source: What it means and how they are different from ChatGPT, Google Gemini in Simple Termsand what it means for users..
Bengaluru-based AI startup Sarvam AI has released two powerful AI models – Sarvam 30B and Sarvam 105B – as open-source, meaning anyone can download, use and build on them for free. Launched at the India AI Impact Summit in New Delhi last month, these models received praise from tech leaders like Google CEO Sundar Pichai. What makes these models special is the fact that both of them have been built entirely from scratch in India, trained on massive datasets curated in-house, using computing power provided under the IndiaAI Mission.Unlike simply adapting a foreign model, Sarvam says it handled every part of the process itself — from data collection and model design to training and deployment. Co-founder Pratyush Kumar called it proof that “India can build state-of-the-art AI models from scratch.”
How Sarvam models serve different purposes
Sarvam 30B is built for speed – just like Gemini Flash. It handles real-time conversations efficiently and is already powering Samvaad, Sarvam’s conversational agent platform. It can process up to 32,000 tokens of text at a time — roughly the length of a short novel.Sarvam 105B is the heavy lifter – like Gemini Pro. It is designed to handle complex, multi-step tasks, it supports a much larger context window of 128,000 tokens and powers Indus, Sarvam’s flagship AI assistant. Both use a mixture-of-experts architecture – a smart design that activates only a portion of the model’s total parameters at any given time, keeping costs low without sacrificing quality.
How Sarvam’s models are compared to ChatGPT and Gemini
Sarvam’s claims are bold and are backed by benchmark numbers. Sarvam 105B scored 96.7 on AIME 2025 (a tough maths competition test) and 71.7 on LiveCodeBench v6 (a coding benchmark), placing it among the top reasoning models in the world. On agentic tasks — where AI must take real-world actions like browsing websites or calling software tools — it scored 68.3 on Tau2, beating OpenAI’s GPT model at 65.8. Perhaps most impressively for Indian users, the model scored 70 out of 75 on JEE Mains 2026 in a single attempt, and a perfect 75 in two attempts.The company has also claimed that Sarvam 105B has also outperformed Google’s Gemini 2.5 Flash and DeepSeek R1 – a 600-billion parameter Chinese model – on Indian language technical benchmarks.
Sarvam AI: Built for India
What truly sets Sarvam apart from global rivals is its deep focus on Indian users because both models support all 22 official Indian languages and are optimised for voice-first interaction. The models have been trained on trillions of tokens of Indian-language data, including Hinglish — the blended Hindi-English commonly used in everyday speech and messaging.