Cura 1T: Specialized Model for Agentic Healthcare
actAVA AI, :, Haolin Chen, Leon Qi, Steve Brown, Deon Metelski, Tao Xia, Joonyul Lee, Qixuan Wang, Kevin Riley, Frank Wang, Weiran Yao
cs.AI
2026-07-16
A 1T-param healthcare LLM (LoRA on Kimi-K2.6) trained via a human-gated loop that converts failed trajectories into targeted data; tops 5 of 6 medical panels while keeping out-of-domain reasoning.
Healthcare asks three very different things of a model at once: patient-facing answers that follow physician guidelines, clinical reasoning over text and images, and multi-turn diagnosis with correct EHR tool calls. These capabilities break in different ways, and the standard fix (dumping more medical data into one generic update) tends to fix one behavior while eroding another. Cura 1T wants a single model that holds all three, without brute-forcing it with general medical corpora.
The core is a human-gated self-evolution loop driven by an LLM training agent, one target capability per round:
The search object is the data recipe, not the hyperparameters. Each round proceeds only after human sign-off, and a round that backfires is reverted. The final model trains on the consolidated mixture accumulated across all completed capability loops. Base model is Kimi-K2.6, 1 trillion parameters, 256K context, LoRA rank 32.
Cura 1T ranks first on five of six healthcare panels:
| Benchmark | Base Kimi-K2.6 | Cura 1T | Best frontier baseline |
| MedAgentBench (EHR tools) | 0.883 | 0.940 | Claude Opus 4.8 0.937 |
| HealthBench Professional | 0.503 | 0.662 | (none reported) |
| HealthBench Hard | 0.222 | 0.368 | (none reported) |
| MedXpertQA overall | 0.569 | 0.655 | GPT-5.5 0.675 |
| AgentClinic overall | 0.754 | 0.796 | Claude Opus 4.8 0.794 |
The one panel it does not lead is MedXpertQA multimodal (0.722 versus GPT-5.5's 0.771). Out-of-domain holds up: AIME, GPQA-Diamond, and τ²-Bench match frontier comparators, and Retail/Telecom exceed publicly reported scores.
The hard part of building a vertical model is not the domain data, it is that every fix breaks something else. Cura 1T's contribution is less the trillion parameters and more the loop that turns failed trajectories back into training signal; it reframes model iteration from "add more medical data" to "fix this specific failure." For teams building specialized models, the human-gated refine loop is a directly portable workflow, and the fact that it runs on LoRA keeps iteration affordable.
The authors are explicit: compute and data constraints cap results at the current training regime; behavior coverage needs broadening, long-horizon agentic work is untouched, and full-parameter updates are left to future resources. More importantly, high benchmark scores do not equal clinical readiness. The paper states plainly that Cura 1T is a research model, not a medical service, and cannot replace a clinician. Whether LoRA rank 32 can support deeper capabilities is an open question. The frontier baselines on the medical panels are matched to publicly reported numbers, so the exact evaluation settings may not align, and the head-to-head should be read with that caveat.