Indian startup Sarvam has released two open source multilingual AI models under Apache 2.0, making them publicly downloadable via AIKosh and Hugging Face as part of India’s push for sovereign AI and Indic-language systems.
Indian AI startup Sarvam has released two foundational multilingual large language models (LLMs) as open-source software under the Apache 2.0 licence, positioning them as India-built alternatives to global AI systems. The models—featuring 30 billion and 105 billion parameters—are available for public download through AIKosh and Hugging Face, and can also be accessed via Sarvam’s Indus AI chatbot and the company’s API developer dashboard.
The models were first unveiled at the India-AI Impact Summit 2026 in New Delhi, with the official rollout announced on March 6. Sarvam has increasingly emerged as a key contributor to India’s ‘sovereign AI’ initiative, which aims to reduce dependence on foreign AI providers such as OpenAI and Anthropic while developing AI tailored for Indian languages and local applications.
Training was supported by the Rs 10,372-crore IndiaAI Mission, with infrastructure from data centre operator Yotta and technical support from Nvidia, using government-backed GPU compute resources.
Both models use a Mixture-of-Experts (MoE) transformer architecture, activating only a fraction of parameters at a time to improve efficiency and reduce compute costs. The Sarvam 30B model features a 32,000-token context window for conversational applications, while the 105B model offers a 128,000-token context window designed for complex reasoning and agentic workflows.
The models incorporate optimisations such as Grouped Query Attention (GQA) in the 30B model and Multi-head Latent Attention (MLA) in the 105B model to reduce memory requirements during inference.
Sarvam also developed a custom tokenizer supporting 22 scheduled Indian languages across 12 scripts, improving token efficiency for Indic text. Benchmark tests show the 105B model competing with pt-oss 120B and Qwen3-Next, while outperforming DeepSeek R1, Gemini 2.5 Flash, and o4-mini on the Tau 2 Bench for agentic reasoning and task completion.
“Building these models required developing end-to-end capability across data, training, inference, and product deployment. With that foundation in place, we are ready to scale to significantly larger and more capable models, including models specialised for coding, agentic, and multimodal conversational tasks,” Sarvam said in a blog post.
