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
fnParam-2-17B

Param-2-17B

Multilingual MoE model for Indian languages and reasoning

BharatGen released Param-2-17B, a Mixture-of-Experts language model with 17 billion total parameters. It activates only 2.4 billion parameters per token to ensure efficient inference while supporting English, Hindi, and 21 other Indian languages.

Share on WhatsApp

  • 17B total parameters with only 2.4B active per token
  • Supports English, Hindi, and 21 Indian languages
  • Trained on approximately 22.5 trillion tokens
  • Hybrid MoE architecture with 64 specialized experts
  • Context window of up to 32,768 tokens

What it is

Param-2-17B is a large language model developed by BharatGen. It uses a Mixture-of-Experts architecture to balance high capacity with computational efficiency. The model contains 17 billion parameters in total but activates only 2.4 billion for each token processed. This design allows it to handle complex tasks while remaining faster than dense models of similar size. It supports English, Hindi, and 21 additional Indian languages, aiming to improve digital inclusion for underrepresented linguistic communities. The model is designed for reasoning, mathematics, code generation, and tool use.

How it came about

BharatGen built the model from scratch with a focus on linguistic diversity and cultural grounding. The training process involved three distinct phases. The first phase used 14.05 trillion tokens, heavily weighted towards multilingual Indian content. The second phase added 8.08 trillion tokens, and a third long-context phase contributed 500 billion tokens. In total, the model saw approximately 22.5 trillion tokens during pretraining. The developers used NVIDIA H100 GPUs for 24 days to complete this extensive training regimen.

Inside the model

The architecture is a Hybrid Mixture of Experts system. It combines shared experts, which are always active, with routed experts that activate selectively. This structure includes 64 specialized experts in total. Two shared routing experts help maintain stable cross-lingual representations, aiding code-switching between the 23 supported languages. The model has 21 hidden layers, uses Grouped Query Attention, and employs RMSNorm for normalization. It features a vocabulary of 128,000 tokens and a maximum context length of 32,768 tokens. A longer variant supporting 128k context is planned for future release.

What it can do

The model demonstrates advanced capabilities in thinking, reasoning, mathematics, and code generation. It supports tool calling, allowing it to interact with external systems. Benchmarks and examples show it can perform chain-of-thought reasoning for mathematical problems. It also generates structured outputs like JSON, Markdown, and tables effectively. The developers recommend using deterministic decoding for consistent results in production environments. Prompt tuning is essential for maximizing accuracy on specific tasks.

How to use it

The model weights are available openly via Hugging Face. Users can download the model and run it locally using standard libraries like Transformers. BharatGen provides a parser to extract reasoning steps, tool calls, and final answers from the output. The company suggests setting sampling to false for reliable performance in evaluations. The model is suitable for further fine-tuning and diverse downstream applications.

Safety and testing

BharatGen emphasizes responsible AI development by reporting the environmental impact of training. The process consumed approximately 258,048 kWh of energy, resulting in an estimated 194 metric tons of CO2 equivalent emissions. This calculation assumes an India national grid average carbon intensity. The team used curated data and quality classifiers for 22 Indian languages to ensure data integrity during preprocessing.

Questions people ask about Param-2-17B

How many parameters does Param-2-17B have?

The model has 17 billion total parameters. However, it uses a Mixture-of-Experts architecture that activates only 2.4 billion parameters for each token, making inference more efficient.

Which languages does this model support?

It supports English, Hindi, and 21 other Indian languages including Bengali, Tamil, Telugu, Marathi, and Gujarati. This broad coverage aims to serve diverse linguistic communities in India.

What is the maximum context length?

The model supports a context length of up to 32,768 tokens. This allows it to process long-form documents and multi-turn conversations effectively. A version with 128k context is planned.

Can the model perform mathematical reasoning?

Yes, the model is designed with advanced reasoning capabilities. It can solve mathematical problems using chain-of-thought processing and generate structured outputs like tables and JSON.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Param-2-17B yet, so the facts on this page come from the sources listed above: huggingface.co/bharatgenai/Param2-17B-A2.4B-Thinking. 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.