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-M

Sarvam-M

Hybrid reasoning model for Indian languages and English

Sarvam-M is a multilingual language model built by Sarvam AI on top of Mistral Small. Released in May 2025, it features hybrid thinking modes for complex reasoning and general chat. The model focuses on improving performance in Indian languages, mathematics, and coding through specialised post-training.

Share on WhatsApp

  • Built on 24 billion parameter Mistral Small model
  • Supports 10 Indian languages in script and romanized forms
  • Features hybrid think and non-think reasoning modes
  • Released with Apache 2.0 open weights in May 2025
  • Shows 86% gain in romanized Indian language math tasks

What it is

Sarvam-M is a text-only language model developed by Sarvam AI. It is built upon the Mistral Small architecture, which contains 24 billion parameters. The model was released in May 2025 with a focus on Indian languages and English. It supports both formal native scripts and romanized transliterations for ten major Indian languages. The system includes a hybrid thinking mode that allows users to switch between deep logical reasoning and efficient general conversation.

How it came about

Sarvam AI announced the model in May 2025 as part of a series of technical releases aimed at building a sovereign AI ecosystem in India. The team started with the open-weight Mistral Small model, which holds an Apache 2.0 licence. They applied extensive post-training techniques to enhance its capabilities. This process involved supervised fine-tuning and reinforcement learning with verifiable rewards. The goal was to create a model that reflects Indian cultural values while improving technical skills in math and code.

Inside the model

The development process began with curating a dataset from over 11.5 million prompts. After deduplication and quality filtering, the team selected 3.7 million high-quality samples. About one-third of these were translated into Indian languages, including Hindi, Bengali, and Tamil. The training used supervised fine-tuning to adjust the model character and reduce bias. Reinforcement learning with verifiable rewards followed, using the GRPO algorithm. The model supports two modes: think mode for complex tasks and non-think mode for casual chat. Inference optimizations include FP8 quantization and lookahead decoding via TensorRT-LLM.

What it can do

The model shows significant gains over its base version. Benchmarks indicate a 20 per cent average improvement on Indian language tests. Math benchmark scores rose by 21.6 per cent, and programming scores increased by 17.6 per cent. Performance in romanized Indian language math tasks improved by 86 per cent. It outperforms Llama-4 Scout in most areas and matches larger dense models like Llama-3.3 70B. However, English knowledge benchmarks such as MMLU saw a slight drop of about one percentage point compared to the baseline.

How to use it

Sarvam-M weights are available for download on Hugging Face under an Apache 2.0 licence. Users can run the model locally using libraries like Transformers or vLLM. The company also provides an API for cloud access. The API supports different reasoning effort levels, allowing developers to enable or disable the thinking mode. Deployment options include configurations for low cost or high throughput on NVIDIA H100 GPUs.

Safety and testing

The training pipeline included steps to debias completions with political slants. The team re-biased the model towards outputs that reflect Indian cultural values. They used custom scoring processes to filter prompt completions from other models. This character training aimed to ensure the model behaves appropriately within its target cultural context. No specific safety incidents or red-teaming results are detailed in the provided materials.

Reception and debate

The release was presented as a technical exploration into post-training optimizations. The company highlighted the model's efficiency and accuracy in Indic tasks. Comparisons were drawn against other open and closed models to demonstrate competitive performance. The blog post invited feedback and suggestions for future collaborations.

Why it matters

Sarvam-M demonstrates how post-training can specialize a general model for regional languages. It provides a reference for building hybrid reasoning systems with open weights. The detailed methodology for data curation and reinforcement learning offers insights for other developers. By supporting romanized inputs, it addresses common usage patterns in India. This approach may influence how future models handle multilingual and transliterated text.

Questions people ask about Sarvam-M

What languages does Sarvam-M support?

The model supports English and ten Indian languages: Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Oriya, Punjabi, Tamil, and Telugu. It handles formal native scripts, code-mixed text, and romanized transliterations.

How does the hybrid thinking mode work?

Users can enable think mode for complex logic, math, or coding tasks. For general conversation, non-think mode offers faster responses. The API allows setting reasoning effort to low, medium, or high to control this behaviour.

Is Sarvam-M free to use?

Yes, the model weights are open source under an Apache 2.0 licence. You can download them from Hugging Face. Sarvam AI also offers a paid API service for those who prefer cloud access without managing infrastructure.

What is the base model for Sarvam-M?

It is built on Mistral Small, a 24 billion parameter decoder-only transformer. Sarvam AI applied supervised fine-tuning and reinforcement learning to this base to improve its performance in Indian languages and reasoning tasks.

Good reading and sources

    Where this page comes from

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