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
wavShuka v1

Shuka v1

Audio-native language model for Indic languages

Shuka v1 is an encoder-decoder model by Sarvam AI that processes spoken audio in Indian languages. It combines the Saaras v1 audio encoder with Meta’s Llama3-8B-Instruct decoder to enable zero-shot question answering across multiple Indic tongues.

Share on WhatsApp

  • Combines Saaras v1 encoder with Llama3-8B-Instruct decoder
  • Trained on less than 100 hours of audio data
  • Supports zero-shot QA in 11 Indic languages and English
  • Only the 60M parameter projector is fine-tuned
  • Available as open weights on Hugging Face

What it is

Shuka v1 is a multimodal language model designed to understand spoken audio natively, with a specific focus on Indic languages. Developed by Sarvam AI, it functions as an encoder-decoder system. The model allows users to ask questions via audio files and receive text responses. It supports a wide range of Indian languages including Hindi, Bengali, Tamil, and Telugu, alongside English. The system is notable for its efficient training process, requiring less than one hundred hours of audio data to achieve performance across these languages.

How it came about

The model was built by Sarvam AI as part of their effort to create frugal yet effective AI solutions for Indian languages. It represents a combination of existing technologies rather than a wholly new architecture from scratch. The developers leveraged their own in-house audio encoder, Saaras v1, and paired it with Meta’s Llama3-8B-Instruct model. This approach allowed them to utilise established strengths in both audio processing and text generation without retraining the entire network.

Inside the model

The architecture connects two distinct components: the Saaras v1 audio encoder and the Llama3-8B-Instruct decoder. A small projector layer with approximately sixty million parameters links these two parts. During the training phase, the weights of the encoder and decoder remain frozen. Only the projector weights are fine-tuned using data in English and Hindi. Despite this limited fine-tuning data, the multilingual capabilities of the Saaras encoder enable the model to perform zero-shot question answering in other Indic languages effectively.

What it can do

Shuka v1 demonstrates strong performance in zero-shot question answering across various Indic languages. Testing has confirmed its ability to handle audio inputs in Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu. The model processes audio samples at a sampling rate of 16000 Hz and generates text responses up to 512 new tokens. Its design allows it to generalise well to languages not explicitly included in the fine-tuning dataset.

How to use it

The model weights are available openly on Hugging Face under the identifier sarvamai/shuka_v1. Users can run the model using the Hugging Face transformers library. The setup requires specific versions of libraries such as transformers, peft, and librosa. The model runs on GPU devices and supports bfloat16 precision. A demo video is provided by the developers to showcase its capabilities, and sample audio files are available for testing the pipeline.

Questions people ask about Shuka v1

What languages does Shuka v1 support?

The model supports English and ten Indic languages: Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu. It performs well in zero-shot scenarios for these languages.

How was Shuka v1 trained?

Training was frugal, using less than 100 hours of audio. Only the small projector connecting the encoder and decoder was fine-tuned on English and Hindi data, while the rest of the network remained frozen.

What is the architecture of Shuka v1?

It is an encoder-decoder model. It uses Sarvam AI’s Saaras v1 as the audio encoder and Meta’s Llama3-8B-Instruct as the text decoder, linked by a projector with about 60 million parameters.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Shuka v1 yet, so the facts on this page come from the sources listed above: huggingface.co/sarvamai/shuka_v1. 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.