Sarvam 30B is an open-weight Mixture-of-Experts language model developed by Sarvam AI, designed for practical deployment in resource-constrained environments while handling multilingual voice calls and tool calls. It has 2.4B non-embedding active parameters within a 32B-total-parameter MoE architecture using 19 layers, top-6 routing, a dense FFN intermediate size of 8192, and an MoE intermediate size of 1024. The model is positioned as a high-performance multilingual LLM optimized for Indian languages, emphasizing strong reasoning, reliable coding ability, and best-in-class conversational quality across 22 Indian languages. It is released under the Apache License and supports inference through Hugging Face, vLLM, and SGLang.
In a comparative benchmarking study, Sarvam 30B demonstrated strong results on foundational coding evaluations, posting 92.1 on HumanEval and 92.7 on MBPP, alongside a BrowseComp score of 35.5 in the agentic AI domain. The same study notes that Sarvam AI reports the broadest benchmark coverage among Indian foundation models, with results spanning general-purpose reasoning, coding, and Indic-language tasks such as IndiVibe and MILU. Sarvam 30B is classified as a Tier 1 fully indigenous model, trained from scratch on internally curated datasets using IndiaAI Mission compute at Yotta's Shakti H100 cluster. The model fits well for developers building India-focused applications that need conversational multilingual quality, instruction following, and tool use at a manageable inference footprint, with the cataloged API limit token context window supporting longer Indic-language interactions.
