Introducing CURA 1T

actAVA's healthcare-specialized model. Strongest on 5 of 6 healthcare benchmark panels, and priced for the way agents actually work.

Today we're launching CURA 1T, our healthcare-specialized language model. It's a 1-trillion-parameter model, post-trained from Kimi-K2.6 through recursive self-improvement, and built for the three things healthcare AI has to do well: talk to patients safely, reason like a clinician, and act inside real EHR systems.

The short version. CURA 1T is the strongest model on 5 of 6 healthcare benchmark panels we ran, it stays competitive with frontier generalists on reasoning outside healthcare, and it runs at 5 to 20 times lower cost per output token than the models it beats.

Why build for the domain

Healthcare doesn't hand over training signal the way coding or math does. Useful supervision sits scattered across clinical guidelines, board exams, medical images, and EHR workflows. Add examples to sharpen one behavior and you can quietly break another that already worked.

General frontier models cover for that with raw scale. We took the other path and trained for the work clinicians and health systems do every day.

What CURA does

Patient care. High-stakes communication like consultation, triage, and safety-conscious guidance. CURA is trained against physician-authored rubrics, so it communicates clearly, surfaces red flags, and escalates when it should. Patient care is where the training loop moved the model furthest.

Clinical reasoning. CURA works board-level cases across 17 medical specialties and 11 body systems, reasoning over clinical images alongside patient records and examination results.

Healthcare agentic workflows. CURA runs multi-turn diagnostic dialogue and real execution. It takes a history, orders tests, narrows the differential, then drives FHIR tool calls against live EHR systems to order labs, place referrals, and record decisions.

The benchmarks

We evaluated CURA 1T on six healthcare panels covering patient-facing response quality, expert reasoning across text and images, and agentic execution against live EHR tooling. CURA leads 5 of the 6 and ranks 2nd on the sixth.

A few results worth calling out. On HealthBench Professional, the physician-rubric panel, CURA reaches 66.2, the strongest score of any model and more than 10 points clear of Claude Opus 4.8. On HealthBench Hard it scores 36.8, a 14.6-point jump over its Kimi-K2.6 base. On AgentClinic it doubles the base model's NEJM case score, from 0.400 to 0.800, and on MedAgentBench it hits 94.0 task success calling tools against a running FHIR server.

Specialization without the usual tax

Specializing a model usually costs you general ability. CURA 1T stays in the top 5 on every out-of-domain leaderboard we track, and leads outright on AIME 2025 (96.7), τ²-Retail (88.6), and τ²-Telecom (100.0). The healthcare focus didn't erode general reasoning or agentic skill, which is the whole point of the retention data we anchor training on.

How we trained it

CURA learns in rounds. Each round, a training agent picks a target capability, trains with SFT, RL, and continual learning, then folds the results back into the next data mixture based on where the model failed.

A human gates every plan and every keep-or-revert call. Reverted rounds stay in the record, including one that lifted headline scores while quietly damaging a held-out subset. We kept that receipt on purpose, because knowing what didn't transfer matters as much as knowing what did.

The economics change the math

Specialization changes the cost structure. Per output token, CURA 1T costs 5x less than Gemini 3.1 Pro, 10x less than Claude Opus 4.8, 12x less than GPT-5.6, and 20x less than Claude Fable 5, the frontier models it's benchmarked against.

Launch pricing is $0.50 per million input tokens, $2.50 output, and $0.20 for cached input, at any context length up to 256K. For agentic healthcare work, where a single task can burn millions of tokens re-reading records and calling tools, that gap compounds on every turn.

Built to drop into your stack

CURA ships with an OpenAI-compatible API. Point any OpenAI SDK at inference.actava.ai/v1, change the model id to actava/cura-soar, and you're running. 256K context, text and vision, function calling, streaming, and prompt caching.