In one of the biggest advancements to the Indian technology community, Bengaluru-based Sarvam AI has now formally published its underlying reasoning models, Sarvam 30B and Sarvam 105B under open source.
These models were announced in the India AI Impact Summit 2026, and are a major milestone of Sovereign AI. These models are not developed on foundations of foreign architectures, as most of those preceding it have done but have been developed in India.
With the weights availed on platforms such as Hugging Face and AIKosh, Sarvam AI is challenging developers to create high-performance cost-effective applications that are specifically meant to serve the Indian context.
The Power of Local Reasoning
Sarvam 30B and Sarvam 105B will fill in an existing gap in the AI industry: the necessity to have models that are cognizant of the Indian language and culture. The global industry giants such as OpenAI and Google have made advances but their systems tend to be trained with data that is oriented toward the West.
Sarvam AI has trained these reasoning models using trillions of tokens containing quality data in all 22 native official languages of India. This makes the AI to be able to cope with the complicated tasks of regional dialects and even Hinglish language in such a way that the key message can be retained.
Sarvam 30B: The Conversational Workhorse
Sarvam 30B model is faster and enables real time interaction. It is now the driving force of Samvaad, which is the conversational agent platform of the company.
-
Context Window: It has a token that is 32,000 in size, which is ideal on prolonged conversations.
-
Architecture: It is based on a Mixture-of-Experts (MoE) design, and only activates 1 billion parameters at once.
-
Efficiency: The Features Grouped Query Attention (GQA) to maintain low memory usage and low response time.
This kind of a model suits well on a startup that seeks a balance between the cost of supercomputers and the expenses that the company can afford.
Sarvam 105B: Built for Complex Tasks
Overall, the Sarvam 105B is the flagship in the area of heavy lifting. It will be used to run the Indus AI assistant and do agentic workflows where the AI will have to perform real-life tasks.
-
Massive Context: A context window of 128,000 tokens enables it to deal with a complete document or a lengthy research article.
-
Performance: On various Indian language benchmarks, it outperforms other major competitors on the rest of the globe such as DeepSeek R1 and Gemini 2.5 Flash.
-
Smart Scaling: It has a Multi-head Latent Attention (MLA) feature that makes it efficient in performing long-context tasks with remarkable memory usage.
Surprisingly, Sarvam 105B in its JEE Mains 2026 testing got full marks of 75 out of 75 in the exams that they conduct, showing their high level of mathematical and logical thinking ability.
Why Open Source Matters for India
In selecting an open-source route by use of the Apache 2.0 license, Sarvam AI is opening the door to high-end reasoning models. This relocation enables Indian companies, as well as government organizations to hosting such models within their premises, which guarantee privacy and safety of the data.
The IndiaAI Mission helped in the development and it offered the required GPU compute power. It is a good sign that India is able to develop Silicon Valley level technology in this public-private partnership.
What This Means for Developers
This is a game-changer to the developer fraternity. No longer do you have to be dependent on the costly APIs with foreign companies. Indian engineers could build with Sarvam 30B and Sarvam 105B:
-
Multilingual Customer Support: Bots are also aware of the local lingo and code-switching.
-
Legal and Gov-Tech Tools: Quality processing of documents in local languages.
-
Educational Assistants: This is a set of tools that assist students in preparing to do competitive exams such as JEE or UPSC.
When such reasoning models achieve a greater perceived presence in the Indian digital environment, they will probably create a new generation of localized innovation. Sarvam AI has already shown a high standard that there is more to AI in India other than consumption, but also creation.
