Sarvam-30B and Sarvam-105B MoE Models Released for Indian Languages

A new pull request adds support for the Sarvam architecture in llama.cpp, introducing two advanced Mixture‑of‑Experts models: Sarvam‑30B and Sarvam‑105B. Sarvam‑30B contains 2.4 B non‑embedding active parameters and is engineered for practical deployment in resource‑constrained environments, offering strong reasoning, coding, and conversational quality across Indian languages. Sarvam‑105B expands to 10.3 B active parameters, excelling in complex reasoning, mathematics, and agentic tasks, and consistently matches or surpasses several closed‑source models on frontier benchmarks. Both models are optimized for multilingual voice calls and tool‑call integration, making them suitable for real‑world applications such as web search and technical troubleshooting. Training focused heavily on Indian contexts, delivering state‑of‑the‑art performance across 22 Indian languages for its size. The models are built to run reliably on modest hardware, with Qwen‑based quantization enabling efficient inference. The pull request also adds the necessary code to load and run these MoE models within the existing llama.cpp framework. Users can now experiment with Sarvam‑30B and Sarvam‑105B for reasoning and coding workloads on local hardware.