Sarvam-105B opens India-built reasoning model

This is more interesting than a routine model-card drop: Sarvam is pitching a full-stack sovereign AI story, not just another checkpoint with borrowed branding.

• –Sarvam-105B uses a 105B-parameter MoE design with 10.3B active parameters, which matters because it targets frontier-style capability without frontier-style inference cost on every token.
• –The official benchmarks put it in the same conversation as GLM-4.5-Air, GPT-OSS-120B, and Qwen3-Next-80B-A3B-Thinking, especially on math, coding, and agentic evaluations like BrowseComp and Tau2.
• –The strongest differentiator is Indian language coverage: Sarvam claims state-of-the-art results across 22 scheduled Indian languages, which gives the release a clearer market angle than most open model launches.
• –The release also shows deployment intent, not just research theater: Sarvam documents local and server inference paths and says the model already powers its Indus assistant for complex reasoning workflows.