In February 2026, a Bengaluru-based AI company did something most observers thought was still years away: it released a 105-billion-parameter language model trained from scratch on Indian languages, and made it free for anyone to use. Sarvam AI's Sarvam-105B is the most capable domestically built AI model India has ever produced — and it arrived at exactly the moment India needed to prove that sovereign AI was more than a political talking point.
What Sarvam-105B Actually Is (And What It Isn't)
Sarvam AI released two models in February 2026: Sarvam 30B and Sarvam 105B. The 105B model uses a Mixture-of-Experts (MoE) architecture — meaning it has 105 billion total parameters but activates only approximately 9–10.3 billion parameters per token during inference. This makes it significantly more compute-efficient to run than a dense 105B model.
The model was trained from scratch — not fine-tuned from Meta's Llama or another foreign base model — on datasets spanning all 22 official Indian languages. It supports advanced reasoning, mathematics, coding, and voice-first interaction. Alongside the two language models, Sarvam released a text-to-speech system, a speech-to-text transcription model, and a vision model for parsing documents in Indian scripts.
Training compute was provided through India's government-backed IndiaAI Mission, with infrastructure from Yotta data centers and Nvidia GPUs. The models are available on Hugging Face under Apache License 2.0 — enterprise use, fine-tuning, and commercial deployment all permitted without licensing fees. As we analyzed in our piece on how Anthropic's export controls exposed India's AI dependency, Sarvam's release became even more significant after the U.S. government restricted Indian access to Claude Fable 5.
How It Compares: Sarvam 105B vs. Global Frontier Models
On Indian-language benchmarks, Sarvam 105B is genuinely competitive. For tasks in Hindi, Tamil, Telugu, Kannada, Bengali, and Marathi, the model outperforms GPT-4o and Gemini 3.5 Pro on several standard evaluation sets, according to the company's benchmarking at the India AI Impact Summit in early 2026.
On general English reasoning and coding benchmarks, the picture is different. Sarvam 105B performs comparably to Llama 3 70B-class models and somewhat below the current generation of frontier models from OpenAI and Anthropic. This is expected — Sarvam was never trying to be GPT-5.
The before/after comparison that matters most: before Sarvam 105B, an Indian company building a customer-facing product in Tamil needed to use a foreign model that understood Tamil poorly. After Sarvam 105B, there is a freely available, Indian-built model trained natively on Tamil and 21 other Indian languages. For the majority of Indian AI applications, this is a genuine practical improvement.
The Government's Role — And Its Limits
The IndiaAI Mission provided the compute infrastructure that made Sarvam 105B possible. Training a 105B+ parameter model from scratch requires compute costing tens of millions of dollars at current GPU pricing. Without government-backed compute access, Sarvam would not have been able to afford training at this scale.
Statista data from early 2026 estimates India's AI market at approximately $6 billion, growing at 25–35 percent annually. The IndiaAI Mission's explicit goal is to position India as a global AI hub by 2030. Sarvam 105B is the first concrete product of that strategy that can be evaluated by anyone, anywhere. The limitation: private AI labs in the U.S. and China don't wait for bureaucratic timelines to scale. India needs both government compute programs and private capital at scale — and private capital for frontier model training in India remains significantly smaller than in comparable AI economies.
What Developers Are Actually Building on Sarvam
In the months following the February 2026 release, Indian developers began adapting Sarvam 105B for specific sectors. Healthcare startups in Tier-2 cities are using the model's voice capabilities to build telemedicine assistants in local languages. Agriculture technology companies are deploying it for farmer advisory services in regional dialects. Several ed-tech platforms have fine-tuned the 30B model for multilingual tutoring. As we've covered in Indian AI startups solving real-world problems at scale, the most durable Indian AI businesses will be built on use-case specificity, not raw benchmark performance.
What This Means for You
If you're an Indian developer or enterprise technology buyer: Sarvam 105B is worth evaluating seriously right now, especially for use cases involving Indian languages, voice interfaces, or document processing in Indian scripts. The Apache 2.0 license means you can deploy it in production without licensing costs, run it on your own infrastructure without API dependency, and fine-tune it on your proprietary data without sharing anything externally. Download the model weights from Hugging Face, run the benchmark comparison against your actual tasks, and make the decision based on your specific use case — not on frontier model marketing from foreign providers.
Frequently Asked Questions (FAQs)
Q: What is Sarvam AI and who built the Sarvam 105B model?
A: Sarvam AI is a Bengaluru-based company that released India's largest domestically built open-source language models in February 2026. The Sarvam 30B and 105B models were trained from scratch using computing resources from India's government-backed IndiaAI Mission, with infrastructure support from Yotta data centers and Nvidia.
Q: Is Sarvam 105B free to use for commercial applications in India?
A: Yes. Sarvam 105B is released under Apache License 2.0, which permits commercial use, fine-tuning, and redistribution. Model weights are available on Hugging Face at sarvamai/sarvam-105b and on AIKosh. There are no per-token API fees if you self-host the model.
Q: Which Indian languages does Sarvam 105B support?
A: Sarvam 105B supports all 22 officially recognized Indian languages including Hindi, Tamil, Telugu, Kannada, Bengali, Marathi, Gujarati, Malayalam, Punjabi, and others. The model was trained natively on these languages, not translated or adapted post-training from an English base model.
Q: How does Sarvam 105B compare to ChatGPT and Claude for Indian users?
A: On Indian-language tasks — particularly regional language understanding, voice interaction, and Indian-script document parsing — Sarvam 105B is competitive with or superior to GPT-4o and Claude Fable 5. On general English reasoning, it performs closer to Llama 3 70B. The performance difference depends heavily on your specific use case.
Sarvam AI's 105B release is the most significant milestone in India's domestic AI journey to date. It doesn't close the gap with OpenAI or Anthropic on global benchmarks — but for the 1.4 billion people whose first language isn't English, it's exactly the right kind of AI to build.
