Archive · 31 Jul 2026
Sarvam AI Unveils Trillion-Parameter Model, Low-Cost Pricing for India
Sarvam AI is a company in India that wants to make a very big computer brain, called a trillion‑parameter model, that can write code, help with security, and do science.
They will train it on a huge set of special computer chips from Nvidia.
The model should be ready in about six months.
They also made a cheaper version of their 105‑billion‑parameter model that costs only $0.80 for every million words, which is much cheaper than other big companies.
They showed new tools that can talk, listen, and see, and they opened an office in San Francisco to get more talent.
They want to keep all the AI work in India so people can use it safely and cheaply.
Sarvam AI plans to build a trillion‑plus parameter foundation model in India, targeting coding, cybersecurity, simulation, and scientific reasoning.
The model will be trained on a 10,000‑GPU Nvidia Blackwell cluster, expected to launch within six months.
Sarvam’s existing 105‑B model is priced at $0.80 per million tokens, roughly five times cheaper than OpenAI’s GPT‑5.4 Mini and more than eleven times cheaper than Google’s Gemini 3.5 Flash.
The company launched new voice, vision, and enterprise AI products, including Sarvam Inference, Sarvam Vision 2.0, Saras V4, Bulbul V4, and AI agents for customer service and coding.
Sarvam opened a San Francisco office, appointed AI researcher Devendra Chaplot as advisor, and aims to host inference services in India to promote token sovereignty.
- Who
- Sarvam AI and its founders Pratyush Kumar, Vivek Raghavan, and advisor Devendra Chaplot
- What
- Announced plans to build a trillion‑plus parameter AI model and launch new AI products and services
- When
- July 30, 2026 at the Epoch 2026 event
- Where
- Bengaluru, India (event) and San Francisco, USA (new office)
- Why
- To create a sovereign AI ecosystem, reduce dependence on foreign models, and offer cheaper AI services
- Company
- Sarvam AI
- Model
- Trillion-parameter foundation model
- Expected launch
- Within six months
- Compute
- 10,000 Nvidia Blackwell GPUs
- Pricing
- $0.80 per million tokens for 105B model
