Cura 1T: Specialized Model for Agentic Healthcare
06:00 · July 20, 2026 · arXiv cs.AI RSS
Healthcare spans high-stakes communication, expert reasoning, and workflow execution, yet specialized LLMs that cover these use cases together remain limited. A healthcare model must handle patient consultation, clinical reasoning over text and images, interactive diagnosis, and electronic health record (EHR) tool use. These capabilities fail in different ways, and a narrow update for one task can degrade another. We present Cura 1T, a healthcare-specialized LLM trained through a human-gated self-evolution loop. In each evolution round, a training agent plans a target capability, trains the model, evaluates benchmark trajectories, and refines the data mixture from observed failures. This data-centered loop improves the model through targeted synthetic and curated examples rather than a single generic medical-data update. Across the healthcare evaluation suite, Cura 1T ranks at or near the top among frontier baselines, while remaining competitive on out-of-domain reasoning and agentic benchmarks.
Summary
Cura 1T is a one-trillion-parameter language model specialized for healthcare, built on the Kimi-K2.6 base through low-rank adaptation. It targets three overlapping but distinct capabilities: patient-facing and clinician-facing responses that follow physician guidelines, expert clinical reasoning over text and images, and multi-turn agentic workflows that include diagnosis, tool use, and electronic health record interactions. These tasks produce different failure modes, such as rubric omissions, brittle reasoning, or malformed tool calls, and a single broad medical-data update can improve one area while eroding performance elsewhere.
To address this, the training process uses a human-gated self-evolution loop. In each round an agent identifies a target capability and acceptance criteria, constructs a focused mixture of synthetic and curated examples, trains a candidate model, runs the relevant benchmarks, inspects the graded trajectories, and revises the data mixture on the basis of observed failures. The loop therefore replaces generic medical-data scaling with iterative, failure-driven data construction. The resulting model records the highest scores on five of the six healthcare benchmark panels reported—MedAgentBench, HealthBench Professional, HealthBench Hard, MedXpertQA, and AgentClinic—while placing second on the remaining multimodal subset of MedXpertQA. Gains over the Kimi-K2.6 base range from roughly four to sixteen percentage points depending on the metric.
The same training regimen leaves out-of-domain performance largely intact. Cura 1T remains competitive on general reasoning, mathematics, and non-healthcare agentic tasks, indicating that the targeted healthcare updates do not produce obvious catastrophic forgetting under the evaluations used. The work therefore demonstrates a practical route to a single model that simultaneously supports consultation, expert reasoning, and workflow execution without requiring separate systems for each use case.
Why it matters
This research is highly relevant for Dutch AI researchers and healthcare institutions developing specialized clinical models. The data-centric, self-evolving training methodology offers a transparent and rigorous approach to building reliable healthcare AI, aligning with EU regulatory standards for clinical deployment.
