Krutrim logo

Krutrim models

Ola Group AI company in Bengaluru that built the Krutrim family of Indian-language text, vision, speech and OCR models and serves models through Krutrim Cloud.

Bengaluru, Indiaolakrutrim.com

Releases
9
Open weights
8
Latest
Mar 11, 2026

March 2026

3 releases
  1. MultimodalOpen weights

    Chitranuvad

    Chitranuvad is a multimodal translation model from Krutrim AI Labs that uses an accompanying image to guide translation from English into Indian languages. It is a fine-tuned version of the Chitrarth-1 vision-language model.

    Modalities
    image, text → text
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights
    • Translates English into Hindi, Bengali and Malayalam, and also handles text-only translation and image captioning
    • Built for the English-to-low-resource multimodal translation shared task at WMT 2024, where Krutrim says it was the winning entry
    • Trained in three stages: image-text feature alignment, multimodal instruction tuning and translation-specific fine-tuning
    • Weights and inference code arrived on Hugging Face and GitHub about a year after Krutrim posted the system description on arXiv in February 2025
    Source: krutrim-ai-labs/Chitranuvad · Hugging Face
  2. MultimodalOpen weights

    Chitrapathak-1

    Chitrapathak-1 is an OCR model from Krutrim AI Labs that transcribes printed document images in ten Indian languages by treating OCR as image-to-text generation. It pairs a CLIP vision encoder with the Krutrim-1 7B language model in a LLaVA-style design.

    Modalities
    image → text
    Parameters
    7B language model plus CLIP ViT-L/14 vision encoder
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights
    • Covers Hindi, Sanskrit, Bengali, Marathi, Tamil, Telugu, Kannada, Malayalam, Punjabi and Odia
    • Splits dense pages into aspect-ratio-aware crops plus a global view before encoding them
    • Trained in two stages: projector-only pretraining, then joint fine-tuning of the projector and language model on OCR data
    • Krutrim's February 2026 arXiv paper on OCR for India reported that its successor, Chitrapathak-2, ran 3 to 6 times faster
    Source: krutrim-ai-labs/Chitrapathak-1 · Hugging Face
  3. MultimodalOpen weights

    Chitrapathak-2

    Chitrapathak-2 is the second model in Krutrim's Chitrapathak OCR series, fine-tuned from Nanonets-OCR2-3B to transcribe Indian-language book pages and documents. It covers ten Indian languages and English.

    Modalities
    image, text → text
    Parameters
    3B decoder plus vision encoder
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights
    • Built by fine-tuning an existing OCR model rather than joining a separate vision encoder and language model, as Chitrapathak-1 did
    • Krutrim's arXiv paper reported a 3 to 6 times speedup over Chitrapathak-1
    • Krutrim reported the top Telugu result in its comparison and second place in the other languages it tested
    • Languages are Hindi, Sanskrit, Bengali, Telugu, Tamil, Marathi, Kannada, Malayalam, Odia, Punjabi and English
    Source: krutrim-ai-labs/Chitrapathak-2 · Hugging Face

February 2025

5 releases
  1. LLMOpen weights

    Krutrim-2

    Krutrim-2 is a 12-billion-parameter multilingual language model from Ola's Krutrim, built on the Mistral NeMo architecture and trained for English and 22 Indian languages. It was released the same day Krutrim introduced Krutrim AI Labs.

    Modalities
    text → text
    Context
    125K context
    Parameters
    12B
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights, API, Consumer app
    • 128K-token context window, with supervised fine-tuning followed by direct preference optimization
    • Krutrim reported top-three results on 5 of 7 tasks in its BharatBench suite against larger open and commercial models
    • Offered at launch in the Krutrim chat app, through the Krutrim Cloud API and as downloadable weights
    • Trained between December 2024 and January 2025 on data running up to April 2024
    Source: krutrim-ai-labs/Krutrim-2-instruct · Hugging Face
  2. SpeechOpen weights

    Dhwani-1

    Dhwani-1 is Krutrim's speech language model built on Krutrim-1, which takes in speech directly instead of relying on a separate speech recognition step. Krutrim open-sourced its speech-to-text translation capability between English and Indian languages.

    Modalities
    audio, text → text
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights, API
    • Follows the SALMONN design: Whisper and BEATs encoders feed the Krutrim LLM through a window-level Q-Former, with LoRA adaptation
    • Translates speech between English and Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Tamil and Telugu
    • Post-trained with methods from Krutrim's IndicST paper at ICASSP 2025 plus in-house translation data
    • Weights were gated behind a click-through form on Hugging Face, with hosted access on Krutrim Cloud
    Source: krutrim-ai-labs/Dhwani · Hugging Face
  3. LLMOpen weights

    Krutrim Translate

    Krutrim Translate is a distilled translation model from Krutrim that builds on IndicTrans2 and raises its context length from 256 to 4,096 tokens.

    Modalities
    text → text
    Context
    4K context
    Parameters
    ~450M
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights, API
    • Distilled to 6 encoder and 3 decoder layers; Krutrim reported at least 4 times lower latency than IndicTrans2 and its distilled versions
    • Translates in both directions between English and Hindi, Bengali, Kannada, Marathi, Malayalam, Gujarati, Punjabi, Telugu and Tamil
    • Trained on AI4Bharat's Bharat Parallel Corpus Collection plus Krutrim's own data
    • Weights were gated behind a click-through form on Hugging Face, with hosted access on Krutrim Cloud
    Source: krutrim-ai-labs/Krutrim-Translate · Hugging Face
  4. MultimodalOpen weights

    Chitrarth-1

    Chitrarth-1 is Krutrim's vision-language model for 10 Indian languages and English, pairing the Krutrim-1 7B language model with a SigLIP image encoder.

    Modalities
    image, text → text
    Parameters
    7.5B
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights, API
    • Covers Hindi, Bengali, Telugu, Tamil, Marathi, Gujarati, Kannada, Malayalam, Odia and Assamese
    • Krutrim reported that it outperformed IDEFICS 2 (7B) and PALO 7B on most academic multimodal benchmarks it tested
    • Introduced with BharatBench, Krutrim's evaluation suite for vision-language models in Indian languages
    • Described in a NeurIPS 2024 workshop paper; weights and code released under the Krutrim Community License
    Source: krutrim-ai-labs/Chitrarth · Hugging Face
  5. EmbeddingOpen weights

    Vyakyarth-1

    Vyakyarth-1 is Krutrim's multilingual sentence embedding model for Indian languages, fine-tuned from the STSB-XLM-R multilingual model with a contrastive objective.

    Modalities
    text → embedding
    Context
    0K context
    Parameters
    270M
    License
    Krutrim Community License Agreement 1.0
    At launch
    Open weights, API
    • Maps sentences to 768-dimensional vectors for semantic search, retrieval, clustering and classification
    • Trained for Hindi, Bengali, Gujarati, Kannada, Malayalam, Marathi, Sanskrit, Tamil and Telugu as well as English
    • Krutrim reported higher FLORES Indic-to-English retrieval accuracy than MuRIL, IndicBERT and jina-embeddings-v3
    • Accepts inputs of up to 128 tokens
    Source: krutrim-ai-labs/Vyakyarth · Hugging Face

December 2023

1 release
  1. LLMProprietary

    Krutrim-1

    Krutrim was the first foundation language model from Ola's AI company of the same name, trained from scratch on more than 2 trillion tokens with a large share of Indian-language data. Krutrim later published it as the 7B Krutrim-1.

    Modalities
    text → text
    Context
    4K context
    Parameters
    7B
    • Unveiled at a livestreamed launch on December 15, 2023, with a waitlist for its chat assistant and for developers
    • At launch Krutrim said it supported all 22 scheduled languages and generated text in Hindi, English and 8 other Indian languages
    • Announced alongside Krutrim Pro, a larger multimodal model planned for later release
    • Weights were released on Hugging Face as Krutrim-1 under the Krutrim Community License in February 2025
    Source: Press release - Ola Krutrim Blog

Cite and reuse this data

These release records are free to reuse under CC BY 4.0, including commercially. Credit The AI Engineering and link to this page's URL.

Dates are the labs' official announcement dates, and every record links to its source. List fields in the CSV are separated by " | ".

Citation

The AI Engineering Editors. "Krutrim models: release history and timeline." The AI Engineering, updated Mar 11, 2026. https://theaiengineering.com/releases/krutrim

HTML link

<a href="https://theaiengineering.com/releases/krutrim">Krutrim models: release history and timeline</a> (The AI Engineering)
Full dataset
1,250 releases from 73 labs
Coverage
to
Updated
License
CC BY 4.0
Publisher
The AI Engineering

Other labs