Vaidya 2.0

Medical language models are usually scored the way medical students are: a stem, five options, one right answer. A real clinical exchange is graded on much more than the answer — whether the response is complete, whether it flags the red flags, whether it asks the follow-up question a clinician would ask, and whether it says any of this in a way a frightened person can act on. Vaidya 2.0 is built around that gap. Rather than optimizing for exam accuracy, it is post-trained with rubric-based reinforcement learning over doctor-annotated synthetic healthcare data, where the reward decomposes a good clinical response into accuracy, completeness, empathy, and context, and scores each separately.

Sitting on top of that is a deep-research augmented reasoning framework: parallel agentic threads search the web during a single response, so guidance can be assembled from current sources rather than recited from whatever was frozen into the weights. The practical consequence is that the same checkpoint handles symptom triage, wellness planning, emergency guidance, longitudinal patient support, and administrative summarization without a task parameter — the request is classified implicitly and routed to the matching capability.

Vaidya is developed by Fractal Analytics, which was selected by the Government of India under the IndiaAI Mission to build a healthcare reasoning model. Version 2.0 was launched at the India AI Impact Summit in New Delhi in February 2026, succeeding the 2024 original that Fractal fine-tuned from open-source base models in the 30–70B range on 650,000+ images and 200,000+ text inputs using both LLM and VLM architectures. Weights are not distributed: the model is served as a fixed checkpoint, Vaidya-v2, behind a commercial API.

#Key Features

• Rubric-based reinforcement learning: Post-training rewards are anchored to physician-annotated rubrics covering the several dimensions on which a clinical response is judged, not to a single correct-answer signal.
• Parallel agentic search: The reasoning framework runs concurrent deep-search threads over the web and reasons on top of what they return.
• Multimodal clinical input: The chat schema accepts multimodal content parts, and requests can attach an image or a PDF — a lab report, a photographed medication strip.
• Multilingual by design: 12 or more Indian languages are supported, including Hindi, Marathi, Tamil, Telugu, and Bengali.
• Drop-in serving interface: An OpenAI/vLLM-compatible Chat Completions endpoint at api.vaidya.ai, so existing OpenAI SDK code works against it unchanged.

#Technical Details

Fractal has not disclosed Vaidya 2.0's parameter count, base model, or architecture, and has published no technical report. What is published is the evaluation. On OpenAI's HealthBench (Hard) — a rubric-graded benchmark of realistic healthcare conversations — Fractal reports a score of 50.1 for Vaidya 2.0, against 46.2 for GPT-5 Thinking, 44.6 for Baichuan-M3-235B, 25.5 for Gemini 3 Pro, and 15.5 for MedGemma 27B. Fractal also reports leading performance on MedExpert, which evaluates expert-level medical reasoning. These figures come from Fractal's own documentation and launch materials; no independent replication has been published. The predecessor's published evaluation was exam-style instead: 83% accuracy on two decades of PG NEET postgraduate medical papers. Access is token-metered at ₹125 per million input tokens and ₹835 per million output tokens on ISO 27001-certified infrastructure, and Fractal states that API queries are never used to train or fine-tune the model.

#Applications

The model's first large public deployment is a WhatsApp health chatbot run with the Brihanmumbai Municipal Corporation, whose network handles more than 16.8 million outpatient visits and generates over three million laboratory reports a year. Citizens send a lab report and receive a plain-language summary in English, Marathi, or another supported language, find nearby municipal facilities, and ask general health questions; the wider pilot adds pre-consultation summaries written into the hospital management system for doctors and an anonymized public-health dashboard for administrators. Beyond that, the API targets digital health apps, wellness and benefits platforms, and clinical workflow tools that need triage or record summarization.

#Impact

Fractal positions Vaidya as the first of a series of verticalized foundation models built under the IndiaAI Mission's sovereign-AI programme, which makes it a concrete test of whether a national AI programme in a low-resource health system produces something clinically useful rather than a demonstration. Its design choices — multilingual coverage of Indian languages, delivery over WhatsApp, a municipal rather than a private-hospital first customer — are aimed at that question. The constraints are real: no weights, no technical report, and benchmark results that rest entirely on the developer's own reporting, which makes independent scrutiny of the headline claim impossible for now. Fractal is also explicit that the model informs rather than diagnoses, and that clinical decisions remain with qualified professionals.