Sarvam AI Blows Past DeepSeek and Gemini with New Reasoning LLMs (February 2026)
- The New Titans: Launch of Sarvam-30B (for real-time conversation) and Sarvam-105B (for complex agentic reasoning).
- Efficiency Breakthrough: The 105B model uses a "Mixture-of-Experts" (MoE) architecture, activating only 9 billion parameters at a time to slash inference costs.
- India-First Performance: Outperforms Google's Gemini Flash on Indian language benchmarks and reportedly beats GPT-120B on the advanced MMLU-Pro reasoning scale.
- Hardware Sovereignty: Trained on 4,096 NVIDIA H100 GPUs via government subsidies, making Sarvam the biggest beneficiary of the IndiaAI Mission.
NEW DELHI — In a massive power move at the India AI Impact Summit 2026, Bengaluru-based Sarvam AI has officially challenged the global AI hierarchy by launching two homegrown "reasoning" models. Keeping up with the latest-ai-news, this development signals a major shift in foundational technology production.
The startup, backed by the Indian government's sovereign AI mission, revealed that its new flagship model not only outclasses Google’s Gemini 2.5 Flash on Indic tasks but effectively matches the reasoning depth of DeepSeek’s R1 while being significantly more efficient.
A Billion Conversations on a Feature Phone
Sarvam AI Co-founder Pratyush Kumar stunned attendees by demonstrating "Vikram," a multilingual AI chatbot powered by the new 30B model that switches effortlessly between Hindi, Marathi, and Punjabi.
Unlike Western models that struggle with "token fertility" (the cost of processing non-English text), Sarvam’s architecture is optimized to run even on low-cost feature phones. The startup also showcased the Sarvam-105B, a behemoth with a 128,000-token context window designed for heavy-duty enterprise work.
During a live demo, the model parsed a complex company balance sheet in real-time, delivering financial insights that rival frontier models like DeepSeek R1, despite being one-sixth the size.
The "Sarvam Kaze" and the Future of Sovereign Tech
Beyond software, Sarvam is expanding into the hardware layer with the upcoming launch of 'Sarvam Kaze' —smart eyewear designed to move AI intelligence from screens to the real world. This hardware-software integration is part of a broader strategy to ensure India’s technological sovereignty.
Partnerships with HMD, Qualcomm, and Bosch were also announced to integrate these India-tuned models into everything from car dashboards to next-gen PCs. By open-sourcing these models, Sarvam is effectively handing Indian developers the keys to build a domestic AI ecosystem that doesn't rely on Silicon Valley APIs.
Why This Is a Global Turning Point
This launch marks the moment India shifted from being an AI talent exporter to a primary producer of foundational technology. With the IndiaAI Mission's Rs 10,000 crore backing, Sarvam has proved that specialized, "sovereign" models can outperform general-purpose global giants by focusing on local linguistic nuance and compute efficiency.
For the first time, a domestic startup has leveled the playing field against Google and OpenAI in the race for the next billion AI users. This homegrown innovation ensures that Indian enterprises can deploy high-reasoning agents without the latency or costs associated with overseas servers.
Frequently Asked Questions (FAQ)
Sarvam AI launched the Sarvam-30B for real-time multilingual conversations and the Sarvam-105B for heavy-duty enterprise reasoning and agentic tasks.
It utilizes a Mixture-of-Experts (MoE) architecture, which activates only 9 billion parameters at any given time, significantly reducing the computational power required for inference.
Vikram is a multilingual AI chatbot powered by the Sarvam-30B model that can switch seamlessly between Indian languages like Hindi, Marathi, and Punjabi.