News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Sarvam 2B

Sarvam 2B

Small language model for English and ten Indian languages.

Sarvam AI built Sarvam 2B, a compact language model designed for English and ten Indic tongues. Pre-trained on two trillion tokens using NVIDIA hardware, it aims to serve users in India with efficient text processing.

Share on WhatsApp

  • Trained on 2 trillion tokens from scratch.
  • Supports 10 Indic languages plus English.
  • Tokenizer has low fertility score for efficiency.
  • Built using NVIDIA NeMo on HGX H100s.

What it is

Sarvam 2B is a small language model created by Sarvam AI. It focuses on generating text in English and ten major Indian languages. The version described here is an early checkpoint, not the final product. It contains two billion parameters and was trained from scratch. The goal is to provide capable performance for regional users while keeping the model size manageable for wider deployment.

How it came about

The team released an early version of the model in mid-2024. This initial checkpoint had completed pre-training but lacked post-training adjustments. Developers shared a notebook to help users start with text completions and supervised fine-tuning. They stated that a final checkpoint would follow, trained on a larger mix of four trillion tokens. That future version would balance English and Indic data equally.

Inside the model

Engineers used the NVIDIA NeMo Framework for training. The work ran on HGX H100 systems within the Yotta Shakti Cloud infrastructure. A key design choice was the tokenizer, built specifically for Indic scripts. It achieves an average fertility score of about two, meaning it uses fewer tokens to represent words compared to other models. This efficiency helps the model process languages like Bengali, Hindi, and Tamil more effectively than general-purpose alternatives.

What it can do

The model supports Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu alongside English. Early demonstrations showed it could complete sentences in these languages, such as identifying historical figures. The developers noted that detailed benchmark scores were not yet available at the time of this release. The focus remains on linguistic coverage and token efficiency rather than raw scale.

How to use it

Weights for this early checkpoint are available openly on Hugging Face. Users can download the model and run it locally or via cloud services. The company provided a Google Colab notebook to facilitate immediate testing. This approach allows researchers and developers to experiment with fine-tuning and integration without waiting for the final commercial release.

Questions people ask about Sarvam 2B

Which languages does Sarvam 2B support?

It handles English and ten Indic languages: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu. The tokenizer is optimized for these specific scripts.

Is this the final version of the model?

No, this is an early checkpoint called v0.5. It has not undergone post-training. The developers plan to release a final version trained on four trillion tokens with equal parts English and Indic data.

What hardware was used to train the model?

The training utilized NVIDIA HGX H100 systems. The process was managed through the NVIDIA NeMo Framework and hosted on the Yotta Shakti Cloud infrastructure.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Sarvam 2B yet, so the facts on this page come from the sources listed above: huggingface.co/sarvamai/sarvam-2b-v0.5. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.