Open Models for Indian Healthcare AI
Eka Care releases purpose-built models for speech, documents and text in Indian clinical settings. All models are in the EkaCare Public Healthcare LLMs collection on Hugging Face and on the India AI AIKosh platform. Try them live at medai.eka.care.Speech
Parrotlet-a 2.5 Pro
- A purpose-built ASR model for medical speech in Indian healthcare settings, transcribing Indian English, Hindi, Marathi, Kannada and Telugu, including the heavily code-mixed speech typical of real consultations.
- Combines a Whisper large-v3 encoder with a MedGemma 4B decoder, further tuned with GRPO on medical conversation data. Best Semantic WER in every language against nine production ASR systems on medical conversation sets.
- AIKosh | Hugging Face | Live demo | Launch blog
Parrotlet-A-EN-5B
- A healthcare-optimized ASR model for transcribing English medical speech in Indian clinical settings, combining a Whisper V3 encoder with a MedGemma3 4B decoder.
- Designed for accurate short-form medical audio transcription, it enables high-fidelity speech-to-text conversion for scribe and voice-enabled healthcare applications.
- AIKosh | Hugging Face | Launch blog
Documents & Vision
Parrotlet-v 2.5 Pro
- A fine-tune of Gemma 4 E2B (~2B effective parameters) for structured, schema-driven extraction from Indian medical records: prescriptions, invoices, lab reports and discharge summaries, plus PII detection.
- Prompted with a document image and a single task tag; returns schema-conformant JSON. Released under Apache 2.0 in bf16.
- AIKosh | Hugging Face | Live demo | Launch blog
Parrotlet-V-Lite-4B
- A purpose-built vision LLM optimized for parsing Indian medical records, including lab reports, digital prescriptions, document classification, and PII extraction.
- Deployable via Hugging Face, it enables structured data extraction from medical images and PDFs for healthcare AI and interoperability workflows.
- AIKosh | Hugging Face | Launch blog
Med Doc Classifier
- A lightweight (~96M parameter) encoder-only model on a SigLIP2-base backbone that sorts a document image into one of 27 classes across two hierarchical levels, and also flags whether it is medical and whether it is handwritten.
- Built by training many specialised models, ensembling them, and distilling the ensemble into one compact student. Released under Apache 2.0.
- AIKosh | Hugging Face | Live demo | Launch blog
Document PII Redactor
- Model weights powering the open-source document-pii-redactor, which detects and redacts PII in both text and document images and is light enough to run on CPU.
- Trained to understand Indian names, documents and contexts; the text model works across Indian languages. The core contribution is the PII token classifier, with OCR as a pluggable front-end.
- AIKosh | Hugging Face | GitHub | Live demo | Launch blog
Embeddings
Parrotlet-E
- A state-of-the-art multilingual embedding model optimized for entity-level medical concept representation across 12 Indic languages and English.
- Fine-tuned on 18M+ SNOMED CT and UMLS-aligned term pairs, it enables robust cross-lingual medical coding, semantic search, and retrieval with benchmark-leading performance on Eka-IndicMTEB.
- AIKosh | Hugging Face | Launch blog
Evaluation
Models are evaluated with KARMA – OpenMedEvalKit, our open-source evaluation framework, on the Eka evaluation datasets. Current results are on each model card and in the launch blogs linked above; the Eka Medical ASR dataset card carries a public leaderboard.Models are released under the license stated on their respective Hugging Face model cards. Some models require accepting the terms on Hugging Face before download.