Foundational Models Launched
Sarvam AI has officially launched two significant foundational multi-lingual AI models, marking a pivotal moment for India's push towards AI independence.
These models, boasting 30 billion and 105 billion parameters respectively, are now accessible under an open-source license, specifically the Apache 2.0 license, for commercial use. Their weights can be downloaded via popular platforms like AIKosh and Hugging Face, democratizing access for developers and researchers. This release follows their debut at the India-AI Impact Summit 2026, where they were first showcased. The development was supported by crucial government initiatives, including compute resources from the Rs 10,372-crore IndiaAI Mission, alongside infrastructure backing from Yotta and technical expertise from Nvidia. Sarvam emphasizes that these models were built from the ground up, trained on extensive, high-quality datasets meticulously curated in-house, underscoring a commitment to indigenous AI development and reducing reliance on foreign technology giants.
Architectural Innovations
Delving into the technical architecture, Sarvam's 30 billion and 105 billion parameter models employ a sophisticated Mixture-of-Experts (MoE) transformer architecture. This design intelligently activates only a portion of the model's parameters for any given task, leading to a substantial reduction in computational expenses during operation. The 30B model is optimized for real-time conversations with a 32,000-token context window, while the more robust 105B model extends this to a 128,000-token window, facilitating complex, multi-step reasoning processes. For enhanced efficiency, the 30B model utilizes Grouped Query Attention (GQA) to minimize KV-cache memory without compromising performance. Meanwhile, the 105B model incorporates DeepSeek-style Multi-head Latent Attention (MLA), further reducing memory demands for processing long contexts. These architectural choices reflect a deep understanding of optimizing performance and resource utilization for advanced AI applications.
Training Data and Languages
The comprehensive training datasets for both Sarvam models encompass a wide array of content, including programming code, general web data, specialized knowledge bases, mathematical texts, and extensive multilingual information. A significant portion of the training budget was dedicated to curating a rich multilingual corpus, focusing on the 10 most widely spoken Indian languages. This deliberate focus on local languages is supported by a custom-built tokenizer, trained from scratch, to ensure efficient tokenization across all 22 scheduled Indian languages and their 12 distinct scripts. Sarvam's tokenizer has demonstrated superior performance, requiring fewer tokens on average to represent words compared to other open-source tokenizers when encoding Indic text. This specialized approach is key to the models' efficacy in understanding and generating content relevant to the Indian linguistic landscape.
Performance Benchmarks
Initial benchmarks indicate that the Sarvam 105B model exhibits superior performance compared to its 30B counterpart, suggesting efficient scaling behavior during training. When evaluated against large language models of comparable size, the 105B model achieved results on par with models like pt-oss 120B and Qwen3-Next (80B) in general capabilities. It also stands out for its strong performance in agentic reasoning and task completion, outperforming DeepSeek R1, Gemini 2.5 Flash, and o4-mini on the Tau 2 Bench. However, the 105B model's capabilities in code generation are noted as lagging behind other models on the SWE-Bench Verified benchmark. The 30B model shows slightly better performance in coding (SWE-Bench Verified) and agentic reasoning (Tau2) when compared to Nemotron 3 Nano 30B, though it performs slightly lower on benchmarks like Live Code Bench v6 and BrowseComp. Notably, Sarvam's 30B model achieves 20-40% higher tokens/sec throughput than Qwen3, attributed to optimizations in its code and kernels.
Safety and Sovereign AI
Sarvam AI has implemented rigorous safety protocols throughout the development of its models. During the supervised fine-tuning phase, both the 30B and 105B models were trained on a dataset designed to address both standard and India-specific risk scenarios. This dataset also incorporated prompts derived from automated red-teaming efforts, including adversarial and jailbreak-style inputs, which were then paired with policy-aligned, safe completions for supervised training. This meticulous approach to safety is crucial, especially in the context of India's 'sovereign AI' initiative. While open-weight models promote accessibility and collaboration, questions have been raised about maintaining sovereignty when models can be freely modified and distributed globally. Sarvam's commitment to developing capable, efficient, and safe AI models tailored for Indian languages and use cases represents a significant stride towards national AI self-reliance.
