Introducing Cura 1T, a specialized model for agentic healthcare.
A one-trillion-parameter model post-trained from Kimi-K2.6 through recursive self-improvement — for patient care, clinical reasoning, health administration, and the long-running agentic workflows payers and providers actually run. Available today through an OpenAI-compatible API.
- Parameters
- 1T
- Context window
- 256K
- Modalities
- Text + vision
- Model id
- actava/cura-soar
- Base model
- Kimi-K2.6
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.
Cura 1T is a one-trillion-parameter model post-trained from Kimi-K2.6 through recursive self-improvement for exactly those three. It is the strongest healthcare LLM we know of today, and it was built to be customized for a single enterprise and owned by it.
Healthcare evaluation panels where Cura 1T is the strongest model of any frontier reference. The sixth is a second place.
Points gained over its base model on the hardest physician-rubric panel — where a wrong answer is rarely the failure, and an omitted one usually is.
Accuracy over that same base on the hardest interactive-diagnosis cases, after one kept round of trajectory training.
A healthcare model 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.
Swipe to see the full diagram
Three capabilities, trained together and measured apart.
Healthcare asks for all three at once, and a model that gains one by losing another has not gained anything. The loop that trained Cura treats that trade as the problem to solve, not a cost of doing business.
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.
See it in practiceClinical 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.
See it in practiceHealthcare 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.
See it in practiceThree capabilities, three walkthroughs.
Benchmarks say how often a model is right. These say what it is like to work with. Each film follows one of the capabilities above through a single piece of healthcare work, at the pace it actually runs.
Patient Care in Practice
A patient conversation Cura carries end to end — history, guidance, and the red flag it refuses to let pass without escalating.
Clinical Reasoning in Practice
A specialty-level case worked through out loud: the images and the record read together, the differential narrowed, the reasoning shown.
Healthcare Workflows in Practice
Cura driving a real workflow as a native tool-caller: reading the system of record, taking the next action, and closing the loop.
The output of one round is the engine of the next.
Recursive self-improvement (RSI) is the idea of an AI system that improves its own capabilities, where each improvement makes it better at making further improvements. That is what makes the process recursive rather than merely iterative.
Cura 1T trains through such a loop. 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.
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. Reverted rounds stay in the record: they show which refinements transferred and which overfit a single benchmark.
Swipe to see the full diagram
Swipe to see the full chart
The recursive-learning record, round by round: each kept intervention lifts its target benchmark, labeled with the retained gain. Dashed red branches are rounds the human gate reverted, including one that raised headline scores while damaging a held-out subset. Scores come from the technical report's detail tables.
Six healthcare panels. Five firsts.
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.
Patient care is where the loop moved the model furthest. On HealthBench Professional, Cura 1T's 66.2 is the strongest score of any frontier model, ahead of Claude Fable 5 (66.0) and more than ten points clear of Claude Opus 4.8 (55.8). On HealthBench Hard, the hardest physician-rubric panel, it scores 36.8: 5.3 points ahead of GPT-5.5 and a +14.6 jump over its Kimi-K2.6 base, whose failures were dominated by omitted rubric points rather than outright errors.
On expert reasoning, Cura 1T leads the MedXpertQA text split at 60.0 (GPT-5.5 59.6, Opus 4.8 56.2) and ranks second to GPT-5.5 on the multimodal split at 72.2. MedXpertQA asks expert-level exam questions across 17 medical specialties; the multimodal cases pair diverse clinical images with patient records and examination results. On agentic execution, Cura 1T edges Opus 4.8 on AgentClinic, 79.6 vs 79.4, and doubles its base's performance on the NEJM cases (0.400 → 0.800). On MedAgentBench it reaches 94.0 task success, within 3.7 points of the best frontier reference, after three kept rounds of tool-use repair.
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, and it leads outright on AIME 2025 (96.7), τ²-Retail (88.6), and τ²-Telecom (100.0). The healthcare specialization does not erode its general reasoning or agentic capability.
| Cura 1T | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | Claude Fable 5 | Kimi-K2.6 | |
|---|---|---|---|---|---|---|
| HealthBench Professional | 66.2 | 55.8 | 51.8 | 43.8 | 66.0 | 50.3 |
| HealthBench Hard | 36.8 | 22.2 | 31.5 | 20.6 | — | 22.2 |
| Cura 1T | Claude Opus 4.8 | GPT-5.5 | Kimi-K2.6 | |
|---|---|---|---|---|
| Text | 60.0 | 56.2 | 59.6 | 48.4 |
| Multimodal | 72.2 | 71.0 | 77.1 | 67.2 |
| Cura 1T | Claude Opus 4.8 | GPT-5.5 | Kimi-K2.6 | |
|---|---|---|---|---|
| MedQA | 87.9 | 84.1 | 83.2 | 86.9 |
| MedQA-Ext | 85.0 | 87.4 | 80.8 | 82.7 |
| NEJM | 80.0 | 80.0 | 46.7 | 40.0 |
| NEJM-Ext | 62.5 | 60.8 | 35.8 | 56.7 |
| Overall | 79.6 | 79.4 | 68.4 | 75.4 |
| Cura 1T | Claude Opus 4.8 | GPT-5.5 | Gemini 3.1 Pro | Kimi-K2.6 | |
|---|---|---|---|---|---|
| MedAgentBench | 94.0 | 93.7 | 89.4 | 91.3 | 84.7 |
| Cura 1T | Best other published | ||
|---|---|---|---|
| AIME 2025 | 96.7 | 96.7 | DeepSeek V3.2 |
| AIME 2026 | 93.3 | 98.3† | Gemini 3.1 Pro |
| GPQA-Diamond | 89.9 | 94.1 | Gemini 3.1 Pro |
| τ²-Airline | 76.0 | 84.0 | Claude Opus 4.5 |
| τ²-Retail | 88.6 | 90.8† | Gemini 3.1 Pro |
| τ²-Telecom | 100.0 | 99.3† | Gemini 3.1 Pro |
Swipe to see every model
† Externally reported (vendor announcements / public leaderboards), not run on our harness; protocols may differ. Hover a value for details.
Evaluation notes
- HealthBench Professional / Hard: Physician-authored rubric scores at T=1.0 on the full sets.
- 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.
- AgentClinic: Pass@1 under the tool-native protocol at T=1.0, with simulated patient and measurement agents.
- MedAgentBench: Task success as a native tool-caller against a running FHIR server; the round-by-round development path uses T=0.6.
- Out-of-domain: Pass@1 at T=1.0; each panel shows that leaderboard's top five.
Cura teaches. The model your organization keeps is the one that learns.
Most organizations rent their intelligence. Every correction an expert makes and every outcome a workflow produces flows back into a vendor's model — the operational knowledge compounds, but not for you. ACTAVA closes the loop differently. Governed production work becomes evidence, evidence becomes evaluations, and evaluations drive recursive learning that produces a specialized model your organization owns, inside your compliance boundary.
We proved the loop on ourselves first: Cura 1T is our own healthcare frontier model, post-trained on that same evaluation corpus and training infrastructure. That is why Cura can act as the teacher for a model trained on your workflows rather than the average internet — and why the evidence that trains it never has to leave your side of the line.
Swipe to see the full diagram
- Zero retention
- Prompt and completion bodies are never persisted. Only operational metadata is kept, which is stricter than the industry default.
- No training on your traffic
- No API traffic is used for model improvement without an explicit written opt-in — a signed agreement, not a toggle buried in a settings page.
- Inside your boundary
- The specialized model and the private learning assets behind it stay under your control, in the compliance boundary they were built in.
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?" }
]
}'- 01Get a key. API keys are issued by ACTAVA: join the waitlist and we'll set you up.
- 02Point your SDK. Any OpenAI SDK works: set
base_url="https://inference.actava.ai/v1"andmodel="actava/cura-soar". 256K context, text + vision, function calling. - 03Read the guides. Streaming, multi-turn chat, vision, tool calls, JSON mode, thinking mode, and prompt caching are covered in the documentation.
Access, data handling, and the line we will not cross.
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.
Build on a healthcare model you can keep.
Cura 1T is available through an OpenAI-compatible API today, and it is the teacher for the specialized models your own production work will train tomorrow. Keys are issued by ACTAVA while access is waitlisted.