Overview
- Sarvam AI unveiled two indigenous LLMs, a 30B model named Vikram for real-time conversations with a 32,000-token window and a 105B model for heavier reasoning with a 128,000-token window, with access offered via API.
- Gnani.ai introduced Vachana STT trained on more than 1 million hours across 1,056 domains, already processing about 10 million calls daily at 95th‑percentile latency of 200 ms, with data stored in Indian data centres.
- Gnani.ai launched Vachana TTS that clones a speaker’s voice in 12 Indic languages from under 10 seconds of audio, available via API or on‑prem for regulated sectors with use cases from emergency alerts to education.
- Clients including the Tata Group and Air India are deploying Gnani’s voice stack, and Sarvam’s models are being rolled out through government partnerships such as UIDAI for voice‑based citizen services.
- IIT Bombay–led BharatGen announced Param2, a 17B Mixture‑of‑Experts multilingual model trained across 22 scheduled languages with NVIDIA, to be released on Hugging Face for open fine‑tuning in domains like governance and healthcare.