Overview
- Sarvam released the models under the Apache 2.0 license with downloadable weights on AI Kosh and Hugging Face, and they are accessible via the Indus app and an API.
- Both models were trained from scratch using GPU resources provided through the IndiaAI Mission with infrastructure from Yotta and technical support from Nvidia.
- The 30B model uses Grouped Query Attention with a 32,000-token context window, the 105B model uses Multi-head Latent Attention with a 128,000-token window, and a new tokenizer targets efficient coverage of 22 Indian languages.
- Sarvam says the 30B model powers its Samvaad conversational platform and the 105B model underpins its Indus assistant for complex reasoning and agentic workflows.
- Early benchmark results reported by Sarvam show the 105B model competing with pt-oss 120B and Qwen3-Next (80B) and outperforming DeepSeek R1, Gemini 2.5 Flash, and o4-mini on Tau 2, while lagging on SWE-Bench Verified; the 30B model shows mixed results versus peers and higher tokens-per-second throughput than Qwen3.