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
wavfnSarvam-30B

Sarvam-30B

Open-weight MoE model for Indian languages and coding

Sarvam-30B is a mixture-of-experts language model released by Sarvam AI. It features 2.4 billion active parameters and supports 22 Indian languages. The model is designed for efficient deployment in resource-limited settings while maintaining strong reasoning and coding capabilities.

Share on WhatsApp

  • Supports 22 Indian languages with state-of-the-art performance
  • Uses only 2.4 billion active parameters per token
  • Licensed under Apache for open commercial and research use
  • Handles multilingual voice calls and tool execution
  • Scores 92.1 on HumanEval coding benchmark

What it is

Sarvam-30B is a large language model built by the Sarvam Foundation Models Team. It uses a mixture-of-experts architecture to balance performance with computational efficiency. The system contains 30 billion total parameters but activates only 2.4 billion during inference. This design allows it to run effectively in environments with limited hardware resources. The model specialises in conversational tasks across 22 Indian languages while also providing robust coding and reasoning skills. It supports multilingual voice interactions and can execute tool calls, making it suitable for practical applications beyond simple text generation.

Inside the model

The model employs a mixture-of-experts structure with 128 distinct experts and one shared expert. It uses top-6 routing to select which experts process each token. The architecture includes 19 transformer layers and grouped key-value attention heads to improve memory efficiency. A high rope theta value of eight million ensures stability for long contexts without requiring additional scaling methods. The router balances load without auxiliary loss, and the intermediate sizes are tuned for throughput. Training focused heavily on Indian linguistic contexts to achieve high performance in regional languages while maintaining general capabilities.

What it can do

Benchmark results show strong performance in mathematics and coding tasks. On Math500, it scored 97.0, and on HumanEval, it achieved 92.1. The model handles complex reasoning tests like GPQA Diamond with a score of 66.5. In agentic benchmarks, it reached 34.0 on SWE Bench Verified and 45.7 on the tau-squared Bench. It demonstrates competitive results in writing tasks and multilingual understanding. The system maintains a context window of 65,536 tokens, allowing it to process substantial amounts of information in a single pass.

How to use it

The weights are available under the Apache License, permitting open use and modification. Developers can download the model from Hugging Face. Inference is supported through standard transformer libraries and requires specific configurations for optimal performance. The model works with SGLang and vLLM, though vLLM support may require installing from a custom fork or applying a hot-patch to the registry. These options allow users to run the model efficiently on compatible hardware.

Questions people ask about Sarvam-30B

What makes Sarvam-30B efficient?

It uses a mixture-of-experts architecture that activates only 2.4 billion parameters out of 30 billion total for each token. This reduces memory usage and computational cost during inference.

Which languages does it support?

The model is designed for 22 Indian languages and English. It achieves high performance in conversational tasks across these regional languages due to focused training data.

Can I use it for coding tasks?

Yes, it shows strong coding abilities with a score of 92.1 on HumanEval and 92.7 on MBPP. It is suitable for generating and debugging code in various programming languages.

How do I run this model?

You can download the weights from Hugging Face. It runs using transformers, SGLang, or vLLM. Note that vLLM may require a custom installation or patch for full support.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Sarvam-30B yet, so the facts on this page come from the sources listed above: huggingface.co/sarvamai/sarvam-30b. 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.