Indian AI startup Sarvam has open-sourced its two reasoning models, Sarvam 30B and Sarvam 105B. These models, trained from scratch in India with government-supported compute resources, excel in Indian languages and perform well on global benchmarks.
The models use a Mixture-of-Experts (MoE) architecture for computational efficiency, with the 30B model designed for practical deployment and the 105B flagship model built for enterprise-grade applications and complex reasoning. Open-sourcing the models aims to foster transparency, lower barriers to AI development, and accelerate innovation.
The release is part of a full-stack AI effort, with the startup having developed the complete training and deployment stack. The models already power the company's in-house offerings: the Samvaad conversational platform and the Indus AI assistant.
Main topics: Sarvam's model release, model specifications and architecture, the significance of open-sourcing, training and development context, and intended applications.
Indian artificial intelligence (AI) startup Sarvam has open-sourced its two reasoning models, Sarvam 30B and Sarvam 105B, co-founder Pratyush Kumar announced on Saturday.
âOpen-sourcing the Sarvam 30B and 105B models! Trained from scratch with all data, model research and inference optimisation done in-house, these models punch above their weight in most global benchmarks plus excel in Indian languages. Get the weights at Hugging Face and AIKosh. Thanks to the good folks at SGLang for day 0 support, vLLM support (is also) coming soon,â Kumar wrote in a post on X.
Open-sourcing a model allows researchers, developers, and companies to access and use the modelâs weights and architecture, enabling greater transparency and collaboration. It lowers the barrier to entry for building AI applications, as users can fine-tune or adapt the model for specific tasks without training one from scratch. This also accelerates innovation by allowing the broader developer and research community to experiment with, improve, and build new tools and applications on top of the model, helping expand the overall AI ecosystem.
Sarvamâs 30 billion parameter model is efficient in reasoning and is built for practical deployment. It uses a Mixture-of-Experts (MoE) architecture that activates only a small fraction of its parameters during generation. Despite having 30 billion total parameters, it uses roughly one billion per token, enabling competitive performance on reasoning and agentic tasks while remaining compute-efficient.
Meanwhile, its flagship 105 billion parameter model is an MoE LLM with 10.3 billion active parameters, designed for enterprise-grade applications and strong performance across Indian languages. It is optimised for complex reasoning tasks, particularly in agentic workflows, mathematics, and coding.
An MoE is a machine learning architecture or a framework where a model is composed of multiple specialised sub-networks called âexperts.â Instead of activating the entire model for every input, a routing mechanism selects only a few relevant experts to process each token or task. This means that while the model may contain many total parameters, only a small subset is active at a time. This makes the model more computationally efficient than dense models of the same size while still benefitting from large overall capacity.
In July last year, co-founder Vivek Raghavan had told ET about the company's plans to open-source its models. The models have been entirely trained in India using the compute capacity provided to the startup under the IndiaAI Mission. Both models were trained from scratch using internally curated datasets across pre-training, supervised fine-tuning, and reinforcement learning. This came amid demands from industry veterans to open-source the model that has been built for India and using taxpayers' money.
ET had reported that the company had received the highest subsidy allocated under the mission so far at Rs 98.68 crore out of the total allocation of Rs 246.71 crore for access to 4,096 Nvidia H100 GPUs for six months.
The release represents a full-stack AI effort. In its blog post, the startup revealed that it has developed the complete training and deployment stack, including tokenisation, model architecture, execution kernels, scheduling systems, and inference infrastructure. This allows the models to run efficiently across a wide range of hardware, from high-end GPUs to consumer devices.
Both models are already being used to build the company's in-house consumer offerings. Sarvam 30B is used to back Samvaad, the companyâs conversational agent platform for enterprises, while Sarvam 105B powers Indus, an AI assistant designed for complex reasoning and agentic workflows, a substitute for ChatGPT and similar generative AI chatbots.
In the pre-training stage, the 30B model was trained on 16 trillion tokens and the 105B model on 12 trillion tokens. The dataset covered web data, code, mathematics, and specialised knowledge sources. A significant portion also included content in the 10 most widely spoken Indian languages.
âOpen-sourcing the Sarvam 30B and 105B models! Trained from scratch with all data, model research and inference optimisation done in-house, these models punch above their weight in most global benchmarks plus excel in Indian languages. Get the weights at Hugging Face and AIKosh. Thanks to the good folks at SGLang for day 0 support, vLLM support (is also) coming soon,â Kumar wrote in a post on X.
Open-sourcing a model allows researchers, developers, and companies to access and use the modelâs weights and architecture, enabling greater transparency and collaboration. It lowers the barrier to entry for building AI applications, as users can fine-tune or adapt the model for specific tasks without training one from scratch. This also accelerates innovation by allowing the broader developer and research community to experiment with, improve, and build new tools and applications on top of the model, helping expand the overall AI ecosystem.
Sarvamâs 30 billion parameter model is efficient in reasoning and is built for practical deployment. It uses a Mixture-of-Experts (MoE) architecture that activates only a small fraction of its parameters during generation. Despite having 30 billion total parameters, it uses roughly one billion per token, enabling competitive performance on reasoning and agentic tasks while remaining compute-efficient.
Meanwhile, its flagship 105 billion parameter model is an MoE LLM with 10.3 billion active parameters, designed for enterprise-grade applications and strong performance across Indian languages. It is optimised for complex reasoning tasks, particularly in agentic workflows, mathematics, and coding.
An MoE is a machine learning architecture or a framework where a model is composed of multiple specialised sub-networks called âexperts.â Instead of activating the entire model for every input, a routing mechanism selects only a few relevant experts to process each token or task. This means that while the model may contain many total parameters, only a small subset is active at a time. This makes the model more computationally efficient than dense models of the same size while still benefitting from large overall capacity.
In July last year, co-founder Vivek Raghavan had told ET about the company's plans to open-source its models. The models have been entirely trained in India using the compute capacity provided to the startup under the IndiaAI Mission. Both models were trained from scratch using internally curated datasets across pre-training, supervised fine-tuning, and reinforcement learning. This came amid demands from industry veterans to open-source the model that has been built for India and using taxpayers' money.
ET had reported that the company had received the highest subsidy allocated under the mission so far at Rs 98.68 crore out of the total allocation of Rs 246.71 crore for access to 4,096 Nvidia H100 GPUs for six months.
The release represents a full-stack AI effort. In its blog post, the startup revealed that it has developed the complete training and deployment stack, including tokenisation, model architecture, execution kernels, scheduling systems, and inference infrastructure. This allows the models to run efficiently across a wide range of hardware, from high-end GPUs to consumer devices.
Both models are already being used to build the company's in-house consumer offerings. Sarvam 30B is used to back Samvaad, the companyâs conversational agent platform for enterprises, while Sarvam 105B powers Indus, an AI assistant designed for complex reasoning and agentic workflows, a substitute for ChatGPT and similar generative AI chatbots.
In the pre-training stage, the 30B model was trained on 16 trillion tokens and the 105B model on 12 trillion tokens. The dataset covered web data, code, mathematics, and specialised knowledge sources. A significant portion also included content in the 10 most widely spoken Indian languages.
Meridiano
