On this picture acquired on Feb. 16, 2026, Prime Minister Narendra Modi throughout the inauguration of India AI Impression Expo, at Bharat Mandapam in New Delhi. Sarvam AI co-founder Pratyush Kumar additionally seen.
| Picture Credit score:
PMO through PTI Picture
Sarvam AI on Tuesday unveiled two new massive language models (LLM) on the India AI Summit — a 30-billion-parameter mannequin and a 150-billion-parameter mannequin — each of which have outperformed comparable models of their respective dimension classes throughout key benchmarks.
In April 2025, the Authorities of India chosen Sarvam below the IndiaAI Mission to develop the nation’s first sovereign massive language mannequin, alongside 11 different startups chosen to advance India’s AI ecosystem.
Manufacturing-ready
The 30B-parameter mannequin is pre-trained on 16 trillion tokens and helps a context size of 32,000 tokens, enabling long-running conversations and agentic workflows. Its comparatively small set of activated parameters retains inference prices low.
Talking on the summit, Pratyush Kumar, the cofounder of Sarvam, defined, “The bigger the mannequin, extra the parameters. Whereas the mannequin is extra succesful, it is usually more durable to coach and longer to run in manufacturing. Nonetheless, a 30 billion parameter mannequin immediately is comparatively small. Ours is a mixture-of-experts (MoE) mannequin and has simply 1 billion activated parameters, which means that in producing each output token, it solely prompts 1B parameters.”
Superior reasoning
Sarvam 30B additionally outperforms friends in its class on environment friendly reasoning, he stated. The mannequin is designed as a real-time conversational engine for manufacturing functions, supporting Indian languages and user-facing dialogue experiences.
Throughout world benchmarks spanning arithmetic, coding, and data duties, Sarvam 30B delivers robust outcomes amongst, both surpassing or stays aggressive with models like OpenAI’s GPT-OSS-20B, Alibaba Cloud’s Qwen3-30B, Mistral-3-2-24B and Google’s Gemma 27B on benchmarks together with Math500 which exams superior arithmetic; HumanEval and MBPP which assess code technology and programming skill; Reside Code Bench v6 which measures real-world coding efficiency; and MMLU and MMLU Professional which consider broad, multidisciplinary data and superior reasoning.
Sarvam 105B, the corporate’s bigger MoE mannequin with 9 billion energetic parameters and a 128,000-token context window, is constructed for advanced reasoning. It delivers robust efficiency throughout math, coding, and Indian languages, helps software program engineering duties reminiscent of bug fixes and code technology, and performs on par with main open- and closed-source frontier models in its class.
On benchmarks, Sarvam performs on par with models reminiscent of GPT-OSS-120B, Qwen3-Subsequent-80B, and Zhipu AI’s GLM-4.5-Air throughout GPQA Diamond, which evaluates graduate-level, multi-step reasoning throughout domains, Past AIME, which exams superior mathematical problem-solving past Olympiad-level issue, and MMLU Professional.
Kumar famous that Sarvam 105B performs higher with Indian languages, in comparison with larger and dearer models like Gemini 2.5 Flash. On most benchmarks, the mannequin additionally beats DeepSeek-R1, a 600B parameter mannequin launched a 12 months in the past. “Whereas these models can evolve quick, Sarvam 105B was skilled from scratch, is one-sixth the dimensions, and but is offering intelligence aggressive to the sooner model of DeepSeek,” he stated.
India-focussed edge
Srinivas Padmanabhuni, the CTO of AiEnsured, a testing suite for AI merchandise, famous that the Sarvam models can doubtlessly broaden entry to underprivileged, colloquial audiences and native use circumstances.
“Within the context of Indian language–based mostly chatbots, whether or not for customer support or queries associated to authorities companies, these will be delivered to bottom-of-the-pyramid customers in native Indian languages. Equally, AI can democratise authorized reasoning and analysis, making them accessible to the bottom-of-the-pyramid and regional language viewers,” he stated.
Sarvam’s models have outperformed bigger ones like Gemini because of stronger dealing with of Indian language context, together with code-mixed codecs like Hinglish, which will be difficult for Gemini Flash models.
Along with focused coaching on native information, Sarvam advantages from optimised architectures reminiscent of mixture-of-experts designs. In the meantime, many world models lack a deep Indian linguistic and cultural context, giving Sarvam a transparent USP in localised language and speech understanding.
All three models, together with the 3B variant, have been skilled utilizing compute allotted below the India AI Mission, with infrastructure help from Yotta and Nvidia.
Sarvam will open-source its 30B and 105B models, enabling builders to construct functions and conversational experiences on high of them.
Revealed on February 18, 2026
Source link
#Sarvam #unveils #indigenouslybuilt #30B #105B #LLM #models
