Krutrim-2
A 12-billion parameter multilingual language model focused on Indian languages and context.
Developed by the OLA Krutrim team, Krutrim-2 is a 12-billion parameter dense transformer released in early 2024. It supports over 128,000 tokens of context and offers strong performance in Indic languages, aiming to match larger models in regional tasks while maintaining competitive English capabilities.
- 12 billion parameter dense transformer model
- Supports context windows up to 128,000 tokens
- Top-3 performance on 5 of 7 BharatBench tasks
- Trained on data up to April 2024
- Open weights released for public use
What it is
Krutrim-2 is a large language model with 12 billion parameters, created by the OLA Krutrim team. Released in January 2024, it is designed to handle both English and a wide range of Indian languages. The model uses a dense transformer architecture derived from Mistral-NeMo. It supports a context window of up to 128,000 tokens, allowing for long documents and extended conversations. Its primary goal is to provide high-quality performance for users in India, focusing on cultural relevance and linguistic nuance.
How it came about
The OLA Krutrim team trained Krutrim-2 between December 2024 and January 2025, according to some records, though release dates are listed as January 2024. The base model and its instruction-tuned variant were both made available on 31 January 2024. This release followed Krutrim-1, a smaller 7-billion parameter model. The developers have stated they continue to improve the system through post-training methods like reinforcement learning from human feedback.
Inside the model
The model is built on the Mistral-NeMo 12B architecture, functioning as an auto-regressive decoder. It uses rotary embeddings with a theta value of one million and a vocabulary size of 131,072 tokens. Training data included web content, code, mathematics, books, and specific Indian context data, with information current up to April 2024. After pretraining, the team applied direct preference optimization to align the model for safety, helpfulness, and reasoning. This process aimed to refine how the model follows instructions across diverse tasks.
What it can do
Krutrim-2 performs competitively on English benchmarks such as MMLU and HumanEval, often matching models five to ten times its size. It shows particular strength in Indic language tasks, achieving top scores in sentiment analysis and paraphrase detection. On BharatBench, a benchmark designed for Indian linguistic nuances, it ranked in the top three for five out of seven tasks. Manual evaluations indicated it scored highest for relevance to Indian cultural contexts compared to other anonymised models.
How to use it
The post-trained versions of Krutrim-2, including the instruct variant, are publicly released with open weights. Developers can download the model from platforms like Hugging Face and run it locally using standard libraries such as Transformers. The team recommends using a temperature setting of 0.3 for optimal performance. The model is also accessible via cloud inference services provided by the developer.
Safety and testing
The training dataset includes internet content, which may contain biases or unsafe material. The developers applied extensive post-training techniques to mitigate these risks, including direct preference optimization. Despite these efforts, the model may still amplify biases, generate inaccurate answers, or produce offensive outputs. Users are advised to maintain human oversight when deploying the model for decision-making or sensitive applications.
Why it matters
Krutrim-2 contributes to the development of natively multilingual models that serve the Indian market. By releasing open weights, it allows researchers and developers to build upon a model optimized for Indic languages. The creation of BharatBench alongside the model highlights a shift towards evaluating AI systems on local cultural and linguistic standards rather than relying solely on translated English benchmarks.
Questions people ask about Krutrim-2
What languages does Krutrim-2 support?
It is natively multilingual, with a strong focus on Indic languages and English. The training data specifically included Indian context and various regional languages to ensure better performance for users in India.
How large is the context window?
The model supports a context length of up to 128,000 tokens. This allows it to process long documents, handle extended multi-turn conversations, and perform tasks like document translation effectively.
Is Krutrim-2 open source?
The post-trained versions of the model are publicly released with open weights. Developers can download and use the model, though the specific licence terms should be checked on the hosting platform.
What architecture is it based on?
It uses a dense transformer decoder architecture based on Mistral-NeMo 12B. It features rotary embeddings and a vocabulary size of 131,072 tokens, designed for auto-regressive language generation.
Good reading and sources
There is no English Wikipedia article on Krutrim-2 yet, so the facts on this page come from the sources listed above: huggingface.co/krutrim-ai-labs/Krutrim-2-instruct. 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.