Sign in

ResearchModels

Cura 1TA healthcare foundation model trained by recursive self-improvement.

A one-trillion-parameter model for patient care, clinical reasoning and the long-running agentic workflows payers and providers actually run, offered through an OpenAI-compatible API.

Why a specialized model

Healthcare does not run on general intelligence. Frontier models are trained toward the average of everything written down. The work in front of your teams asks for three specific things at once: patient communication a clinician would sign, expert reasoning over clinical text and images, and reliable execution against the systems payers and providers actually run. A model that is merely good at all three in a demo is not the same as a model built for them.

A healthcare model also has to reach the systems, not just talk about them. Cura is a native tool-caller. On the payer side that means core administrative processing, care management, network management, policy management, and CRM. On the provider side, EHRs over FHIR, revenue cycle and practice management, ERP and finance, scheduling and workforce, and pharmacy. One model, both halves of the stack, in a 256K context window that holds the case and the policy at the same time.

Payer systems

  • Core administrative processing
  • Care management
  • Network management
  • Policy management
  • CRM

Cura 1T

Native tool calls

Provider systems

  • EHRs over FHIR
  • Revenue cycle and practice management
  • ERP and finance
  • Scheduling and workforce
  • Pharmacy
Figure 1Where Cura 1T executes: the payer and provider systems it calls as a native tool-caller.

Three capabilities

Cura 1T was post-trained for exactly those three, together, and each is measured on its own. A model that gains one by losing another has not gained anything. Each film below follows one capability through a single piece of healthcare work, at the pace it actually runs.

Patient care

High-stakes patient communication: consultation, triage, and safety-conscious guidance. Trained with physician-authored rubrics, Cura communicates clearly, surfaces red flags, escalates when it should, and leads the frontier on physician-graded evaluations.

Clinical reasoning

Expert-level medical reasoning over clinical images. Cura works through specialty-board-level cases across 17 medical specialties and 11 body systems, and reasons natively over diverse clinical images, patient records, and examination results.

Healthcare agentic workflows

Interactive diagnosis and EHR workflow. Cura conducts multi-turn diagnostic dialogues (taking history, ordering tests, narrowing the differential) and drives FHIR tool calls against live EHR systems for providers and core administrative processing systems for payers.

Figure 2The three capabilities. Select a film to play it.

Recursive self-improvement

Cura 1T trains through a recursive self-improvement loop, in which the output of one round is the engine of the next (our survey covers the field). In each round, a training agent selects a target capability and trains the model using SFT, RL, and a continual learning method inspired by self-distillation (SDFT): we add a hint, let the model roll out the task on-policy with it, then train it to execute without the hint. The loop grades the benchmark trajectories, reads the failures, and refines the next data mixture from what it finds.

  1. 1PlanSelect a target capability
  2. 2TrainSFT, RL and hint-based self-distillation
  3. 3EvaluateGrade the benchmark trajectories
  4. 4RefineRead the failures, adjust the data mixture
  5. 5ReviewKeep the round, or revert it

A kept round becomes the starting point of the next.

Figure 3One round of the loop.

Building a model like this is as much a data-construction problem as a training problem. Healthcare offers thinner training signal than, say, a model for coding or math. Useful supervision sits scattered across guidelines, exams, images, and EHR and admin workflows. And adding examples for one behavior can erode another the model already got right. The loop's job is to find the missing data in the recipe, then prove, benchmark by benchmark, that each addition transfers without unnecessary forgetting.

MedAgentBench

88.3 → 94.0 +5.7

859095100Base model (Kimi-K2.6): 88.3BaseRound 1, Tool-use: 94.3 (kept)R1Round 2, Tool-use + retention: 96.7 (kept)R2Round 3, Harness bug fix: 97.3 (kept)R3Cura 1T, Consolidated data mixture: 94.094.0Release

HealthBench Hard

22.2 → 36.8 +14.6

2025303540Base model (Kimi-K2.6): 22.2BaseRound 1, Behavior correction: 33.2 (reverted)R1Round 2, Clean behavior mix: 37.2 (kept)R2Cura 1T, Consolidated data mixture: 36.836.8Release

MedXpertQA

56.9 → 65.5 +8.6

40455055606570Base model (Kimi-K2.6): 56.9BaseRound 1, Reasoning correction: 54.1 (reverted)R1Round 1L, Reasoning correction, extended: 54.5 (reverted)R1LRound 2, Knowledge injection + retention: 60.3 (kept)R2Round 3, Mixture refinement: 63.6 (kept)R3Round 4, Data mixture tuning: 44.0 (reverted)R4Cura 1T, Consolidated data mixture: 65.565.5Release

AgentClinic

75.4 → 79.6 +4.2

70758085Base model (Kimi-K2.6): 75.4BaseRound 2, Interactive trajectory + retention: 80.7 (kept)R2Cura 1T, Consolidated data mixture: 79.679.6Release
Base model Kept round Reverted round Release (consolidated mixture)
Figure 4Each benchmark from the base model to the release, with the gain at top right. The release consolidates the kept rounds into one data mixture, so it can sit a little below the best round. Reverted rounds stay in the record, including the HealthBench round that raised headline scores while damaging a held-out subset. MedAgentBench rounds were scored on the development path at T=0.6. Hover a point for the change it tested.

Results

We evaluated Cura 1T on six healthcare panels spanning patient-facing response quality, expert clinical reasoning across text and images, and agentic execution against live EHR tooling. Cura 1T is the strongest model on five of the six panels and ranks second on the remaining one, MedXpertQA multimodal.

HealthBench Hard

  • Cura 1T36.8
  • GPT-5.531.5
  • Claude Opus 4.822.2
  • Kimi-K2.622.2
  • Gemini 3.1 Pro20.6

HealthBench Professional

  • Cura 1T66.2
  • Claude Fable 566.0
  • Claude Opus 4.855.8
  • GPT-5.551.8
  • Kimi-K2.650.3
  • Gemini 3.1 Pro43.8

MedXpertQA Text

  • Cura 1T60.0
  • GPT-5.559.6
  • Claude Opus 4.856.2
  • Kimi-K2.649.3

MedXpertQA Multimodal

  • GPT-5.577.1
  • Cura 1T72.2
  • Claude Opus 4.871.0
  • Kimi-K2.666.9

AgentClinic

  • Cura 1T79.6
  • Claude Opus 4.879.4
  • Kimi-K2.675.4
  • GPT-5.568.4

MedAgentBench

  • Cura 1T94.0
  • Claude Opus 4.893.7
  • Gemini 3.1 Pro91.3
  • GPT-5.589.4
  • Kimi-K2.684.7
Figure 5Cura 1T (rose) against frontier references on the six healthcare panels: physician-rubric patient care (HealthBench Professional and HealthBench Hard), expert clinical reasoning (MedXpertQA text and multimodal), and agentic execution (AgentClinic, MedAgentBench). Each panel is sorted by score; models without a reported score are omitted. Protocols are in the footnotes.

Patient care is where the loop moved the model furthest: on HealthBench Hard it gained 14.6 points over its Kimi-K2.6 base, whose failures were mostly omitted rubric points rather than outright errors. On AgentClinic's NEJM cases it doubled the base's accuracy (0.400 → 0.800), and on MedAgentBench it reached the top score after three kept rounds of tool-use repair.

Out of domain

Specialization usually costs generality; retention-anchor data is the loop's counter. Cura 1T stays in the top five on every out-of-domain leaderboard we track, so the healthcare specialization does not erode its general reasoning or agentic capability.

AIME 2025

  • Cura 1T96.7
  • DeepSeek V3.296.7
  • gpt-oss 120B93.4
  • Nova 2.0 Pro89.0
  • Claude Haiku 4.583.7

AIME 2026

  • Kimi-K2.696.4
  • Qwen3.6 Plus95.3
  • MAI-Thinking-194.5
  • Seed 2.0 Pro94.2
  • Cura 1T93.3

GPQA-Diamond

  • Gemini 3.1 Pro94.1
  • GPT-5.593.5
  • Claude Opus 4.892.0
  • Kimi-K2.691.1
  • Cura 1T89.9

τ²-Airline

  • Claude Opus 4.584.0
  • Qwen3.581.5
  • Gemini 3 Pro80.5
  • Cura 1T76.0
  • Claude Sonnet 4.572.0

τ²-Retail

  • Cura 1T88.6
  • Claude Sonnet 4.586.2
  • Gemini 3 Pro85.3
  • Qwen3.584.4
  • DeepSeek V3.281.1

τ²-Telecom

  • Cura 1T100.0
  • Qwen3-Max98.2
  • Claude Sonnet 4.598.0
  • Gemini 3 Pro98.0
  • Qwen3.597.8
Figure 6Each panel shows its leaderboard's top five for agentic tool use (τ²-Airline, τ²-Retail, τ²-Telecom), competition math (AIME 2025 and 2026), or graduate-level science QA (GPQA-Diamond). Competitors vary by panel; Cura 1T (rose) ranks in all six.

Using Cura 1T

Change the base URL and the model id; nothing else. Point any OpenAI SDK at inference.actava.ai/v1 and swap the model id. Two endpoints are served: /v1/models and /v1/chat/completions.

curl https://inference.actava.ai/v1/chat/completions \
  -H "Authorization: Bearer $ACTAVA_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "actava/cura-soar",
    "messages": [
      { "role": "user", "content": "Escalation criteria for chest pain triage?" }
    ]
  }'

Streaming, multi-turn chat, vision, tool calls, JSON mode, thinking mode and prompt caching are covered in the documentation: Quickstart · API reference · GitHub

Cura 1T is a research model, not a medical service, and not a substitute for a clinician. Benchmark scores do not establish safety for unsupervised clinical use.

Footnotes

  1. 01

    HealthBench Professional / Hard: Physician-authored rubric scores at T=1.0 on the full sets.

  2. 02

    MedXpertQA: Expert-level medical exam QA spanning 17 specialties and 11 body systems; exact-letter pass@1 at T=1.0; the overall score combines 2,450 text and 2,000 multimodal questions.

  3. 03

    AgentClinic: Pass@1 under the tool-native protocol at T=1.0, with simulated patient and measurement agents.

  4. 04

    MedAgentBench: Task success as a native tool-caller against a running FHIR server; the round-by-round development path uses T=0.6.

  5. 05

    Out-of-domain: Pass@1 at T=1.0; each panel shows that leaderboard's top five.

BenchmarksCHI-Bench: Can AI Agents Automate End-to-End, Long-Horizon, Policy-Rich Healthcare Workflows?LearningRecursive Self-Improvement in AI: A Survey
All research

Build on a healthcare model you can keep.

Keys are issued by ACTAVA while access is waitlisted.