LearningUnit 2.0 (L1-30B-A5B)

L1-30B-A5B is the Korean-locale medical foundation model from Lunit and Lunit Consortium. It is the 30B member of the L1 family, post-trained directly from Gravity-30B-A5B-Base, a sparse Mixture-of-Experts model developed by Trillion Labs and the Lunit Consortium.

L1-30B-A5B preserves the base model's GravityMoEForCausalLM architecture. GravityMoE is weight- and attention-compatible with DeepSeek-V3 while retaining its own Hugging Face architecture name and remote-code registration. Load the model with trust_remote_code=True.

• 🇰🇷 Korean clinical locale, trained natively — not machine-translated from English
• ⚡ 5.34B active of 29.56B total (sparse MoE)
• 💭 Reasons in <think>...</think> before answering — budget 2048+ output tokens
• 🔧 Multi-turn retrieval and tool-call trajectories with citation-grounded answers
• 📏 131,072-token context

📊 Benchmark

All numbers were produced with CoEval, Lunit's open-source medical LLM evaluation framework.

Knowledge & reasoning

Chat quality

[!NOTE] HealthBench rubric items are graded by an LLM judge: zai-org/GLM-5.2-FP8, thinking disabled (chat_template_kwargs.enable_thinking: false). GLM was chosen because it grades more strictly than the GPT-family judges and keeps a GPT model from judging its own family.

🚀 Quickstart

⚠️ Required settings

repetition_penalty = 1.05 — at the default 1.0 the model degenerates into repeated phrases on long clinical answers. It ships in generation_config.json, but any client that sets its own sampling parameters overrides that. It is not an OpenAI API parameter — through the OpenAI SDK it must go in extra_body or it is dropped silently.

System prompt — the identity line the model was aligned on, and the one every benchmark number below used. Task-specific prompts go after it, not instead of it.

client.chat.completions.create(
    model="learning-unit/L1-30B-A5B",
    messages=[{"role": "system", "content": "You are Chain-of-Evidence, a medical AI assistant developed by Lunit."}, *messages],
    temperature=0.0,          # 0.0 for benchmarks and deterministic clinical tasks
    top_p=1.0,
    max_tokens=32768,
    extra_body={"repetition_penalty": 1.05},
)

SGLang

Native GravityMoE support is being upstreamed to SGLang. Until that work is merged, use Trillion Labs' sglang-gravity fork, which registers GravityMoEForCausalLM on SGLang's DeepSeek-V3-compatible implementation.

pip install "sglang[all] @ git+https://github.com/trillion-labs/sglang-gravity.git#subdirectory=python"

Launch the server with the same GravityMoE deployment path as the base model, substituting the L1 checkpoint:

python3 -m sglang.launch_server \
  --model-path learning-unit/L1-30B-A5B \
  --host 0.0.0.0 --port 30000 \
  --tp 4 --context-length 131072 \
  --trust-remote-code --dtype bfloat16

Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

name = "learning-unit/L1-30B-A5B"
model = AutoModelForCausalLM.from_pretrained(name, torch_dtype=torch.bfloat16,
                                             device_map="auto", trust_remote_code=True)
tok = AutoTokenizer.from_pretrained(name, trust_remote_code=True)

messages = [
    {"role": "system", "content": "You are Chain-of-Evidence, a medical AI assistant developed by Lunit."},
    {"role": "user", "content": "성인 환자에서 Obstructive sleep apnea가 의심될 때, 진단을 위해 어떤 검사를 시행하고 어느 전문과로 의뢰하는 것이 적합한가요?"},
]
inputs = tok([tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)],
             return_tensors="pt").to(model.device)
out = model.generate(**inputs, max_new_tokens=2048, do_sample=True, temperature=0.7,
                     repetition_penalty=1.05)
print(tok.decode(out[0][inputs.input_ids.shape[1]:], skip_special_tokens=True))

🔧 Tool use

Tool calling was trained end-to-end against Lunit's own RAG harness — the Chain-of-Evidence MCP servers, with tool schemas frozen at a snapshot. Two consequences:

• Handed a tool that looks nothing like a retrieval tool, the model often answers from parametric knowledge instead of calling it. Force it with tool_choice: "required", or keep your schema close to what it knows: a query string plus optional filters, returning ranked passages with stable ids.
• Citation markers index into what the retrieval tool returned that turn. With no retrieval tool attached they have nothing to point at.

💬 Examples

Representative prompts per capability. 🔧 expects a retrieval tool call first.

🇰🇷 Korean clinical practice

• 국시원 국가고시 — 65세 남성, 3개월간 기침과 체중 감소, 30갑년 흡연력, 흉부 X선상 우상엽 3 cm 결절. 가장 먼저 시행할 검사는? ① 흉부 CT ② 객담 세포검사 ③ 기관지내시경 ④ PET-CT ⑤ 경피적 세침흡인생검
• KTAS 응급 분류 — 이 환자의 KTAS 중증도(1–5)를 분류하세요. [환자] 78세 여성, 22시경 갑자기 좌측 안면마비와 구음장애, 두통 호소, 고혈압 병력, BP 172/88, HR 92, SpO₂ 98%, 의식 명료.
• 응급실 감별진단 — 54세 남성, 2시간 전 시작된 상복부 통증과 식은땀, 30갑년 흡연·당뇨, BP 148/92 HR 104, 심전도 미확인. 놓치면 안 되는 진단부터 우선순위로.
• 청구 관련 업무 🔧 — 만성 특발성 두드러기에 오말리주맙 급여 인정기준과 인정 투여 기간을 근거 고시와 함께. / 프레가발린 75 mg 처방에 상병이 M54.5 하나인 청구 건, 누락된 인정상병이 있는지.
• 의료 법령 🔧 — 진료기록부 보존기간과, 전자의무기록을 외부 클라우드에 보관할 때의 요건을 근거 조문과 함께.

🧠 Clinical reasoning

• Differential diagnosis — 45F with lupus nephritis on mycophenolate and prednisone, fever, dry cough, bilateral ground-glass opacities, CD4 180. Differential and workup?
• ADR causality — 80세 여성, moxifloxacin 400 mg IV 투여 중 전신 소양감 발생, 중단 후 호전, 재투여 없음. WHO-UMC 기준으로 인과관계 평가.
• Clinical calculation — 72세 남성, 68 kg, Cr 1.4 mg/dL. Cockcroft-Gault 로 CrCl 계산 후 아픽사반 용량 조정 필요 여부.

📚 Evidence & communication

• Literature Q&A 🔧 — Does perioperative continuation of SGLT2 inhibitors increase euglycemic DKA risk? Cite every claim.
• Guideline Q&A 🔧 — 2형 당뇨에서 SGLT2 억제제를 1차 약제로 고려할 수 있는 조건을 지침 근거와 함께.
• Patient education — I use insulin daily. How should I store it at home?
• Clinical documentation — Overnight progress note from labs, vitals, and a stable nursing report.

⚠️ Limitations

• Not a substitute for professional medical judgment. Outputs may be wrong, incomplete, or outdated, and need review by a qualified clinician.
• Knowledge is frozen at the training cutoff — guidelines, 고시, and drug approvals move. Korean reimbursement answers in particular must be re-verified against current notices.
• Thinking costs tokens. <think> reasoning raises latency and token use.
• Tool calling is harness-shaped. Quality degrades the further your tool schemas drift from the ones it was trained on. See Tool use above.

📝 Citation

@misc{lunit2026l1_30b,
  title={L1-30B-A5B: A Korean-Locale Clinical Language Model by Lunit},
  author={Lunit},
  year={2026},
  url={https://huggingface.co/learning-unit/L1-30B-A5B}
}

🤝 Acknowledgements

This work was supported by the Domain-Specific Foundation Model Project (인공지능 특화 파운데이션 모델 프로젝트), funded by the Ministry of Science and ICT (과학기술정보통신부) and managed by the National IT Industry Promotion Agency (NIPA).

L1-30B-A5B is a collaborative effort by the following consortium members:

Industry

• Lunit
• Trillion Labs
• SK Biopharmaceuticals
• Kakao Healthcare
• AIGEN Sciences
• D-Circle
• Rebellions
• Standigm

Academia

• Prof. Choi Yun-jae's Lab from KAIST
• Prof. Hong Seung-hoon's Lab from KAIST
• Prof. Jung Yu-seong's Lab from SNU
• Prof. Kim Hyun-woo's Lab from KAIST
• Prof. Kim Tae-gyun's Lab from KAIST
• Prof. Ye Jong-cheol's Lab from KAIST

Hospitals

• NHIS Ilsan Hospital
• Ewha Womans University Seoul Hospital
• Keimyung University Dongsan Medical Center
• Konyang University Hospital
• Korea University Research & Business Foundation
• Kyung Hee University Hospital at Gangdong
• Kyung Hee University Medical Center
• Pusan National University Yangsan Hospital
• Yongin Severance Hospital

📄 License

This model is licensed under the Apache 2.0 License.

📬 Contact

• Taesoo Kim (김태수) — taesoo.kim@lunit.io
• Donggeun Yoo (유동근) — dgyoo@lunit.io