Cura 1T

🌐 Learn more at actava.ai/cura.
🔑 Get an API key: join the waitlist.
📍 Use the Cura API — API documentation (OpenAI-compatible).
📖 Check out the Cura 1T Technical report.

Introduction

Cura 1T is actAVA's one-trillion-parameter healthcare model. Fine-tuned from Kimi-K2.6 through recursive self-improvement — each round, a training agent plans a target capability, trains the model, evaluates benchmark trajectories, and refines the data mixture from observed failures, with humans gating every keep-or-revert decision — it handles the work healthcare models are actually asked to do:

• Patient consultation: open-ended health conversations, graded hard by physician-written rubrics
• Clinical reasoning over text and medical images: expert-exam-level diagnosis, treatment, and basic-science questions (text + vision, 256K context)
• Interactive diagnosis and EHR tool use: multi-turn consultations and FHIR-based record operations, end to end

Cura 1T leads frontier models on 5 of 6 healthcare benchmark panels: HealthBench Hard (36.8 vs 31.5 for GPT-5.5), HealthBench Professional (66.2, edging Claude Fable 5's 66.0), MedXpertQA-Text (60.0), AgentClinic (79.6 vs 79.4 for Claude Opus 4.8), and MedAgentBench (94.0 vs 93.7 for Claude Opus 4.8); GPT-5.5 keeps the lead on MedXpertQA-Multimodal (77.1 vs 72.2). Against its base model Kimi-K2.6 the gains are wide: +14.6 points on HealthBench Hard, +15.9 on HealthBench Professional, +8.6 on MedXpertQA overall, +9.3 on MedAgentBench, and +4.2 on AgentClinic.

Specialization does not cost general capability: Cura 1T stays frontier-competitive out of domain — 96.7 on AIME 2025, 89.9 on GPQA-Diamond, and top scores on τ²-bench Retail (88.6) and Telecom (100.0).

For more detail, check the technical report.

Reported scores use the tech report's protocol — pass@1 with sampling (temperature 1.0 for HealthBench, MedXpertQA, and AgentClinic; MedXpertQA overall = 2,450 text + 2,000 multimodal questions). cura-eval below defaults to greedy decoding for reproducible measurement — pass --temperature 1.0 to match the report's setting.

API

The Cura API is OpenAI-compatible: base URL https://inference.actava.ai/v1, model id actava/cura-soar, bearer auth with ACTAVA_API_KEY (join the waitlist for a key).

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": "Summarize the escalation criteria for chest pain triage."}]
  }'

Existing OpenAI SDK code works by changing the base URL and model id — see the API documentation for streaming, vision, tool calling, and thinking mode. Ready-made configs for local agents (Claude Code, Hermes, OpenClaw, Cline, Roo Code) are in examples/integrations/.

Evaluation with cura-eval

This repo contains cura-eval, the medical evaluation harness used to evaluate Cura 1T. It is model-agnostic: every benchmark drives the model under test through a plain OpenAI-compatible chat-completions endpoint, so the same commands evaluate Cura, OpenAI, OpenRouter models, or anything you serve locally with vLLM / SGLang / Ollama / LM Studio.

Benchmark	Command	Grading	Needs
HealthBench (hard / regular / consensus)	cura-eval healthbench	LLM rubric judge	judge key
HealthBench Professional	cura-eval healthbench-professional	LLM rubric judge + length adjustment	judge key
MedXpertQA — Text	cura-eval medxpertqa --subset text	exact letter match (deterministic)	—
MedXpertQA — MM	cura-eval medxpertqa --subset mm	exact letter match (deterministic)	vision endpoint
AgentClinic (MedQA / NEJM)	cura-eval agentclinic	simulated consultation, moderator-graded	judge key (NPCs)
MedAgentBench	cura-eval medagentbench	agentic FHIR tasks, deterministic graders	FHIR docker server

Protocol notes. MedAgentBench runs on native OpenAI function calling (fhir_get / fhir_post / finish) rather than upstream's strict GET/POST/FINISH text protocol, so scores are not directly comparable to the published leaderboard — it measures native tool-calling agentic ability on the same tasks and graders. AgentClinic follows the upstream "prompt" protocol and grading control flow.

Install

pip install -e .            # or: uv pip install -e .
pip install -e ".[images]"  # + Pillow, for NEJM image downscaling in agentclinic

Python ≥ 3.10.

Quickstart — evaluate Cura

export ACTAVA_API_KEY=...   # candidate model
export OPENAI_API_KEY=...   # judge for the rubric-graded benchmarks

# Rubric-judged health conversations (10-example smoke)
cura-eval healthbench --limit 10

# Board-exam MCQ, judge-free
cura-eval medxpertqa --subset text --limit 50

# Multimodal MCQ (Cura is text + vision)
cura-eval medxpertqa --subset mm --limit 20

# Simulated clinical consultations
cura-eval agentclinic --dataset MedQA --limit 10

# Agentic FHIR tasks (start the server first)
docker run -p 8080:8080 jyxsu6/medagentbench:latest
cura-eval medagentbench --limit 10

Defaults target the Cura API: base URL https://inference.actava.ai/v1, model actava/cura-soar, bearer key from ACTAVA_API_KEY.

Evaluate any OpenAI-compatible endpoint

The candidate model is just --base-url + --model + --api-key-env:

# Local vLLM
cura-eval medxpertqa --base-url http://localhost:8000/v1 \
    --model meta-llama/Llama-3.3-70B-Instruct --api-key-env VLLM_API_KEY

# Ollama
cura-eval healthbench --base-url http://localhost:11434/v1 \
    --model llama3.1 --api-key-env OLLAMA_API_KEY --limit 20

# OpenRouter
cura-eval healthbench-professional --base-url https://openrouter.ai/api/v1 \
    --model moonshotai/kimi-k2.6 --api-key-env OPENROUTER_API_KEY

# OpenAI
cura-eval agentclinic --base-url https://api.openai.com/v1 \
    --model gpt-5.5 --api-key-env OPENAI_API_KEY --temperature none

Notes:

• For keyless local servers, set the key env var to any placeholder (e.g. EMPTY).
• --temperature none omits the field for reasoning endpoints that reject sampling params.
• --extra-body '<json>' forwards provider-specific fields verbatim (OpenRouter provider/reasoning routing, Cura thinking, …).
• The judge is equally model-agnostic: --judge-model, --judge-base-url, --judge-api-key-env. Defaults: gpt-4.1-2025-04-14 (HealthBench), gpt-5.4 (HealthBench Professional), gpt-5.5 (AgentClinic NPCs + moderator) on the official OpenAI API. Keep the judge fixed when comparing candidate models.

Output

Each run writes a fresh directory under --output-dir (default results/):

• responses.jsonl — one row per example: {<id>, "messages": [...], "response": ..., "score": {...}}. messages is the full conversation including the generated assistant turn, with any chain-of-thought lifted into reasoning_content — graded runs double as SFT-ready thinking/tool-calling data.
• summary.json — headline score, per-category breakdowns, and the exact run config.

Headline metrics: HealthBench = mean rubric reward (0–1); HealthBench Professional = mean length-adjusted reward (raw also reported); MedXpertQA = exact-match accuracy; AgentClinic = diagnostic accuracy; MedAgentBench = task success rate.

Python API

import asyncio
from cura_eval import ChatClient
from cura_eval.benchmarks.medxpertqa import run_medxpertqa

candidate = ChatClient(
    model="actava/cura-soar",
    base_url="https://inference.actava.ai/v1",
    api_key_env="ACTAVA_API_KEY",
)
rows, summary = asyncio.run(run_medxpertqa(candidate, subset="text", limit=20))
print(summary)

Repo layout

src/cura_eval/
  client.py                 # the OpenAI-compatible seam (ChatClient)
  loaders.py                # HealthBench / HBP / MedXpertQA dataset loaders (HF)
  judge.py                  # rubric judge (simple-evals template) + scoring
  mcq.py                    # deterministic MCQ letter extraction
  rows.py                   # per-example row schema + run writers
  cli.py                    # `cura-eval` CLI
  benchmarks/
    healthbench.py healthbench_professional.py medxpertqa.py
    medagentbench/          # tools, FHIR client, vendored tasks + graders
    agentclinic/            # doctor, NPCs, episode loop, scenario fetch
examples/                   # endpoint recipes + agent integration configs
tests/                      # offline unit tests (no network, no keys)

Attribution

The evaluation harness ports and builds on:

• HealthBench rubric judge template and scoring — openai/simple-evals (MIT)
• HealthBench Professional — openai/healthbench-professional
• MedXpertQA — TsinghuaC3I/MedXpertQA (Zuo et al., 2025)
• AgentClinic — SamuelSchmidgall/AgentClinic (MIT; Schmidgall et al., 2024). NPC prompts and episode control flow are upstream's; scenario data is fetched from the upstream repo at run time.
• MedAgentBench — stanfordmlgroup/MedAgentBench (Jiang et al., 2025). Task data and reference-solution graders are vendored; see the protocol note above.

Benchmark datasets remain under their upstream licenses/terms; this repo's code is licensed under Apache 2.0 (see LICENSE and NOTICE).

Disclaimer

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.

Citation

If you use Cura 1T or this evaluation harness, please cite:

@misc{ai2026cura1tspecializedmodel,
      title={Cura 1T: Specialized Model for Agentic Healthcare}, 
      author={actAVA AI and : and Haolin Chen and Leon Qi and Steve Brown and Deon Metelski and Tao Xia and Joonyul Lee and Qixuan Wang and Kevin Riley and Frank Wang and Weiran Yao},
      year={2026},
      eprint={2607.15314},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2607.15314}, 
}