Sarvam 105B

Sarvam 105B is Sarvam AI's flagship open-weights reasoning model, released alongside the smaller Sarvam 30B on March 6, 2026. According to the official announcement, it was trained from scratch on large-scale, in-house curated datasets spanning pre-training, supervised fine-tuning, and reinforcement learning, with all training compute provided through the IndiaAI mission. The release covered the full stack, including tokenization, model architecture, execution kernels, scheduling, and inference systems, so the same weights can be deployed efficiently across flagship GPUs and consumer laptops.

Sarvam positions the 105B variant for complex reasoning, programming, and agentic workflows, where it currently powers the Indus assistant. The company reports strong performance on reasoning, code, and agentic benchmarks, and state-of-the-art results on Indian language evaluations relative to larger models. Because the weights are openly available, teams can self-host for private or regulated deployments, while hosted access through Sarvam remains an option for those who prefer an API, making the model a flexible choice for multilingual assistants, tool-using agents, and India-focused applications.