Sarvam AI Releases 105B Parameter Multilingual Model with Support for 23 Indian Languages
Key Takeaways
- ▸Sarvam AI released Sarvam-105B, a 105 billion parameter open-source model under Apache 2.0 license
- ▸The model supports 23 Indian languages plus English, addressing a major gap in multilingual AI capabilities
- ▸Features include tool calling, reasoning mode with 'thinking' tokens, and multi-turn conversation support
Summary
Indian AI startup Sarvam AI has released Sarvam-105B, a large language model with 105 billion parameters optimized for Indian languages. The model has been made available on Hugging Face under an Apache 2.0 license, indicating an open-source release that allows commercial use. The model supports 23 Indian languages including Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese, Urdu, Sanskrit, Nepali, Sindhi, Konkani, Maithili, Dogri, Manipuri, Santali, Kashmiri, and Tibetan, alongside English.
The model appears to implement a custom architecture called 'sarvam_mla' and includes advanced features such as tool calling capabilities and a reasoning mode indicated by 'thinking' tokens in its chat template. The chat template structure suggests the model can handle multi-turn conversations, system prompts, and tool integrations, positioning it as a capable assistant for complex workflows. The model uses the Transformers library and is distributed in the SafeTensors format for efficient loading and deployment.
Since its release on March 3, 2026, the model has garnered 88 likes on Hugging Face, indicating early community interest. The release represents a significant contribution to multilingual AI, particularly for underserved Indian language markets where high-quality large language models have been limited. By open-sourcing a model of this scale with comprehensive Indian language support, Sarvam AI is addressing a critical gap in the AI landscape and enabling developers across the region to build language-specific applications.
- Available on Hugging Face with custom 'sarvam_mla' architecture and SafeTensors format for efficient deployment
Editorial Opinion
Sarvam-105B represents a watershed moment for AI accessibility in South Asia, where linguistic diversity has long been a barrier to AI adoption. The decision to release such a large model under an open-source license is particularly commendable, as it enables local developers, researchers, and businesses to build applications without prohibitive API costs. However, the true test will be the model's performance quality across all 23 languages—maintaining consistency across such linguistic diversity at this parameter scale is technically challenging, and independent benchmarks will be crucial to validate its practical utility.
