India’s First AI: Sarvam-M Launches скачать в хорошем качестве

India’s First AI: Sarvam-M Launches

India has launched its first homegrown foundational AI model. Bengaluru-based Sarvam AI has been selected by the government to build the country’s national AI infrastructure, starting with Sarvam-M—a 24 billion parameter hybrid model built on top of Mistral Small. Sarvam-M is post-trained in Indian languages alongside English, and supports both native scripts and romanized inputs. It offers two processing modes—"think" and "non-think"—and outperforms similar-sized models in math, coding, and colloquial understanding. It also handles informal Indian language and cultural references with high accuracy. A key focus is localization and debiasing. Where previous models gave answers tailored to foreign contexts, Sarvam-M now adapts to Indian ones. For example, its explanation of the Speaker of the House now refers to the Lok Sabha with culturally relevant framing. To measure its performance, Sarvam created the Indic Vibe Check—a benchmark designed to test how well models understand Indian conversational style. In these evaluations, Sarvam-M outperformed Mistral Small, Gemma 3, Llama 4 Scout, and Llama 3.3 across language, math, and programming tasks. Sarvam AI also previously launched Bulbul v2, a text-to-speech model supporting 11 Indian languages with fast, low-latency responses and localized voice tones. It supports brand voices and has been positioned as a speech AI solution built for Indian use cases. Sarvam-M is now available via the Sarvam API playground or as a downloadable model on Hugging Face.