Descriptions:
Fahd Mirza walks through the hands-on local installation and testing of Sarvam-30B, an open-source large language model developed by Indian AI startup Sarvam AI under India’s government-backed AI Mission. The model is one of two releases — a 30 billion parameter version and a larger 105 billion parameter sibling — and represents India’s entry into sovereign foundation model development alongside efforts from the US, China, Europe, and the UAE.
What distinguishes Sarvam-30B is its explicit focus on all 22 scheduled Indian languages, including Hindi, Bengali, Tamil, Telugu, Malayalam, Marathi, Gujarati, Punjabi, Urdu, Sanskrit, and others, supporting both native scripts and the romanized Latin text commonly used in messaging. The architecture uses a Mixture of Experts design with 128 experts, activating only a small subset per token to keep inference fast and memory-efficient. Running on an Nvidia H100 with 80GB VRAM, the model loads at approximately 129GB on disk and consumes around 62GB of VRAM during inference.
Mirza tests the model with a demanding creative prompt — generating a self-contained HTML file displaying culturally relevant phrases or proverbs in all 22 Indian languages, styled around Indian festivals like Holi, Diwali, and Eid. Benchmark comparisons show the model is competitive with similarly-sized models on math, coding, and reasoning, with standout performance on Indian language tasks where it outpaces larger competitors. The video is a practical first-look for developers interested in multilingual open-weight models.
📺 Source: Fahd Mirza · Published March 07, 2026
🏷️ Format: Hands On Build
