Open Source AI Explained
The concept of open-sourcing artificial intelligence models signifies a pivotal shift towards greater accessibility and collaborative development within
the tech community. When models like Sarvam's are made open source, their fundamental code, architectural blueprints, and often the fully trained parameters are released to the public without any licensing fees or restrictions. This transparency invites a broad spectrum of users – from independent researchers and ambitious startups to established businesses – to freely download, examine, and adapt these advanced AI systems for their specific needs and projects. It fosters an environment where the inner workings of AI can be thoroughly understood, leading to more effective debugging and continuous refinement. Furthermore, it cultivates a global ecosystem of developers who can collectively contribute to enhancing the models, rectifying any flaws, and introducing innovative new functionalities, thereby accelerating the overall pace of AI advancement and application development.
Sarvam 30B & 105B: Key Features
Sarvam AI's newly released 30B and 105B models represent significant advancements, distinguished by their creation from the ground up within India and their specialized focus on Indian languages and regional use cases. The '30B' and '105B' designations refer to the approximate number of parameters within each model, with 30 billion and 105 billion respectively; larger parameter counts generally correlate with enhanced capabilities but also demand greater computational resources for operation. A standout characteristic is their profound proficiency in all 22 official Indian languages, alongside Hinglish, the prevalent blended Hindi-English dialect. This deep linguistic integration is crucial for ensuring broad adoption and utility across India. Unlike proprietary systems such as OpenAI's ChatGPT or Google Gemini, which are typically accessed via controlled APIs, Sarvam's open-source ethos actively encourages widespread community engagement, experimentation, and extensive customization, paving the way for tailored AI solutions.
Performance Benchmarks & Comparisons
Sarvam AI's models are not just positioned for local relevance but also demonstrate impressive performance metrics that rival global leaders. The Sarvam 105B model, for instance, achieved a score of 96.7 on the challenging AIME 2025 mathematics competition benchmark and 71.7 on the LiveCodeBench v6 for coding, placing it among the world's top-tier reasoning models. In agentic tasks, which involve AI performing real-world actions like web browsing or interacting with software tools, it scored 68.3 on Tau2, surpassing OpenAI's GPT model at 65.8. Particularly noteworthy for Indian students, the model achieved a remarkable 70 out of 75 on the JEE Mains 2026 exam in a single attempt and a perfect 75 in two attempts. The company further claims that Sarvam 105B has outperformed Google's Gemini 2.5 Flash and China's 600-billion parameter DeepSeek R1 model on specific Indian language technical benchmarks, underscoring its strength in localized AI capabilities.
Architectural Innovations & Use Cases
Both Sarvam 30B and 105B employ a sophisticated mixture-of-experts (MoE) architecture. This design is particularly efficient as it only activates a subset of the model's total parameters for any given task, thereby optimizing resource utilization and keeping operational costs down without compromising quality. Sarvam 30B is engineered for speed, mirroring the performance of models like Gemini Flash, and is adept at handling real-time conversational interactions, currently powering Sarvam's Samvaad conversational agent platform. It boasts a substantial context window capable of processing up to 32,000 tokens, equivalent to a short novel. Sarvam 105B, comparable to Gemini Pro, is designed for more complex, multi-step tasks. It supports a significantly larger context window of 128,000 tokens and is the driving force behind Indus, Sarvam's flagship AI assistant, highlighting their versatility for diverse applications from customer service bots to sophisticated research tools.
Impact of Open Source AI
The decision by Sarvam AI to open-source its 30B and 105B models heralds a transformative era for artificial intelligence, especially within India. This strategic move is instrumental in democratizing access to cutting-edge AI technology, significantly lowering the barrier to entry for smaller businesses, independent developers, and academic institutions. By making these powerful tools freely available, Sarvam AI is poised to catalyze a surge in innovation, encouraging the development of novel AI applications specifically tailored to address local needs and challenges. This increased accessibility also fosters healthy competition within the AI sector, potentially driving down service costs and elevating the overall quality and diversity of AI-powered solutions. Furthermore, the open-source nature offers greater control over data privacy and security, as models can be deployed on-premises, directly addressing growing concerns about data handling and confidentiality in the digital age.
