How to use from the
Use from the
Transformers library
# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="learning-unit/L1-30B-A5B", trust_remote_code=True)
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)
# pip install -U transformers accelerate
# Load model directly
from transformers import AutoModelForCausalLM
model = AutoModelForCausalLM.from_pretrained("learning-unit/L1-30B-A5B", trust_remote_code=True, device_map="auto")
Quick Links

L1-30B-A5B

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.

  • ๐Ÿ‡ฐ๐Ÿ‡ท Korean clinical locale, trained natively โ€” not machine-translated from English
  • โšก 5.34B active of 29.56B total (sparse MoE)
  • ๐Ÿ’ญ Reasons in <think>...</think> before answering โ€” budget 2048+ output tokens
  • ๐Ÿ”ง Multi-turn retrieval and tool-call trajectories with citation-grounded answers
  • ๐Ÿ“ 131,072-token context
Architecture GravityMoE (sparse MoE + MLA) Layers 52 (2 dense, 50 MoE)
Total / active params 29.56B / 5.34B Hidden size 2048
Routed experts 64, top-8 MoE intermediate 1408
Shared experts 1 Attention / KV heads 16 / 16
Context 131,072 Vocab 151,552
Tokenizer GLM-4.5 Precision bf16

๐Ÿ“Š Benchmark

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

Chat quality vs MCQA, bubble area = active parameters

Knowledge & reasoning

Model W.Avg PubMedQA AttrBench MedQA CareQA HeadQA MedMCQA MMLU-Pro (Health) M-ARC MedExQA MetaMedQA MedHallu MedCalc KorMedMCQA MedBullets 4-opt MedBullets 5-opt MedXpertQA-R MedXpertQA-U
L1-30B-A5B 83.76 82.20 76.31 92.30 92.55 89.66 76.45 79.34 42.00 86.17 86.23 82.71 81.55 89.53 84.74 79.87 43.10 37.86
GPT-OSS-120B 79.99 78.00 76.10 91.10 91.00 88.40 74.80 74.60 40.00 84.10 76.50 83.50 30.30 84.80 84.70 82.10 35.60 32.90
GPT-5.6-SOL 87.85 78.60 74.95 96.00 95.60 92.70 84.10 83.10 74.00 87.10 81.40 94.80 80.09 97.71 91.60 87.70 58.10 56.70
GPT-5.6-TERRA 86.97 78.20 75.44 93.80 95.60 92.50 83.00 80.70 72.00 87.00 81.90 92.90 82.55 94.18 89.90 85.10 52.20 50.90
GPT-5.6-LUNA 85.36 76.00 75.21 94.20 94.90 91.20 81.80 80.10 68.00 85.40 79.80 89.00 79.27 93.79 89.00 84.10 45.10 45.80
KIMI-K2.6 84.51 80.20 71.84 94.11 92.07 90.64 82.00 81.91 74.00 85.64 82.45 76.37 61.55 93.52 86.69 82.14 48.52 48.22
QWEN3.6-35B-A3B 81.06 78.80 75.90 89.32 90.55 87.87 76.14 77.87 55.00 82.98 76.47 77.50 70.82 86.34 81.17 78.90 39.49 36.50
QWEN3.8-27B 82.82 77.00 71.24 93.72 92.19 89.66 76.45 80.81 67.00 85.11 79.97 86.19 63.00 87.80 87.34 83.12 43.58 39.90
QWEN3.5-122B-A10B 76.07 76.40 55.68 87.80 86.40 84.00 74.40 73.00 59.00 80.20 73.90 37.50 75.82 85.60 79.20 79.50 35.90 35.30
DEEPSEEK-V4-PRO-0813 86.39 76.60 75.92 94.97 95.48 92.61 83.43 82.52 80.00 87.55 79.68 72.14 55.09 95.68 88.31 86.04 49.87 48.05
MEDGEMMA-27B 72.74 73.40 74.80 84.40 85.00 83.80 71.90 73.00 48.00 80.30 69.60 81.40 24.10 56.40 73.70 68.80 19.10 20.50
GEMMA4-31B 80.25 77.60 76.98 87.40 87.90 85.90 73.00 77.90 79.00 82.00 76.80 79.60 53.10 86.10 76.00 74.44 54.60 48.90
GEMMA4-26B-A4B 75.86 76.40 72.00 81.80 84.50 82.30 67.30 73.50 67.00 78.30 71.50 86.50 45.60 80.70 73.70 67.50 45.10 39.20

Chat quality

HealthBench Main / Professional / Consensus

Model HealthBench Main HealthBench Professional HealthBench Consensus
L1-30B-A5B 52.81 40.01 90.15
GPT-OSS-120B 50.49 23.77 78.72
GPT-5.6-SOL 56.59 63.18 84.98
GPT-5.6-TERRA 52.78 59.36 84.55
GPT-5.6-LUNA 50.73 55.07 83.60
KIMI-K2.6 56.71 43.68 90.68
QWEN3.6-35B-A3B 55.39 41.91 88.29
QWEN3.8-27B 52.71 31.42 84.78
QWEN3.5-122B-A10B 51.28 43.55 86.88
DEEPSEEK-V4-PRO-0813 51.23 33.99 87.43
MEDGEMMA-27B 43.36 11.93 82.18
GEMMA4-31B 46.72 31.92 87.60
GEMMA4-26B-A4B 46.66 31.19 86.85

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.

๐Ÿš€ Quickstart

โš ๏ธ Required settings

repetition_penalty = 1.05 โ€” 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 โ€” through the OpenAI SDK it must go in extra_body or it is dropped silently.

System prompt โ€” 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": "์„ฑ์ธ ํ™˜์ž์—์„œ Obstructive sleep apnea๊ฐ€ ์˜์‹ฌ๋  ๋•Œ, ์ง„๋‹จ์„ ์œ„ํ•ด ์–ด๋–ค ๊ฒ€์‚ฌ๋ฅผ ์‹œํ–‰ํ•˜๊ณ  ์–ด๋А ์ „๋ฌธ๊ณผ๋กœ ์˜๋ขฐํ•˜๋Š” ๊ฒƒ์ด ์ ํ•ฉํ•œ๊ฐ€์š”?"},
]
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))

๐Ÿ”ง Tool use

Tool calling was trained end-to-end against Lunit's own RAG harness โ€” 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.

๐Ÿ’ฌ Examples

Representative prompts per capability. ๐Ÿ”ง expects a retrieval tool call first.

๐Ÿ‡ฐ๐Ÿ‡ท Korean clinical practice

  • ๊ตญ์‹œ์› ๊ตญ๊ฐ€๊ณ ์‹œ โ€” 65์„ธ ๋‚จ์„ฑ, 3๊ฐœ์›”๊ฐ„ ๊ธฐ์นจ๊ณผ ์ฒด์ค‘ ๊ฐ์†Œ, 30๊ฐ‘๋…„ ํก์—ฐ๋ ฅ, ํ‰๋ถ€ X์„ ์ƒ ์šฐ์ƒ์—ฝ 3 cm ๊ฒฐ์ ˆ. ๊ฐ€์žฅ ๋จผ์ € ์‹œํ–‰ํ•  ๊ฒ€์‚ฌ๋Š”? โ‘  ํ‰๋ถ€ CT โ‘ก ๊ฐ๋‹ด ์„ธํฌ๊ฒ€์‚ฌ โ‘ข ๊ธฐ๊ด€์ง€๋‚ด์‹œ๊ฒฝ โ‘ฃ PET-CT โ‘ค ๊ฒฝํ”ผ์  ์„ธ์นจํก์ธ์ƒ๊ฒ€
  • KTAS ์‘๊ธ‰ ๋ถ„๋ฅ˜ โ€” ์ด ํ™˜์ž์˜ KTAS ์ค‘์ฆ๋„(1โ€“5)๋ฅผ ๋ถ„๋ฅ˜ํ•˜์„ธ์š”. [ํ™˜์ž] 78์„ธ ์—ฌ์„ฑ, 22์‹œ๊ฒฝ ๊ฐ‘์ž๊ธฐ ์ขŒ์ธก ์•ˆ๋ฉด๋งˆ๋น„์™€ ๊ตฌ์Œ์žฅ์• , ๋‘ํ†ต ํ˜ธ์†Œ, ๊ณ ํ˜ˆ์•• ๋ณ‘๋ ฅ, BP 172/88, HR 92, SpOโ‚‚ 98%, ์˜์‹ ๋ช…๋ฃŒ.
  • ์‘๊ธ‰์‹ค ๊ฐ๋ณ„์ง„๋‹จ โ€” 54์„ธ ๋‚จ์„ฑ, 2์‹œ๊ฐ„ ์ „ ์‹œ์ž‘๋œ ์ƒ๋ณต๋ถ€ ํ†ต์ฆ๊ณผ ์‹์€๋•€, 30๊ฐ‘๋…„ ํก์—ฐยท๋‹น๋‡จ, BP 148/92 HR 104, ์‹ฌ์ „๋„ ๋ฏธํ™•์ธ. ๋†“์น˜๋ฉด ์•ˆ ๋˜๋Š” ์ง„๋‹จ๋ถ€ํ„ฐ ์šฐ์„ ์ˆœ์œ„๋กœ.
  • ์ฒญ๊ตฌ ๊ด€๋ จ ์—…๋ฌด ๐Ÿ”ง โ€” ๋งŒ์„ฑ ํŠน๋ฐœ์„ฑ ๋‘๋“œ๋Ÿฌ๊ธฐ์— ์˜ค๋ง๋ฆฌ์ฃผ๋ง™ ๊ธ‰์—ฌ ์ธ์ •๊ธฐ์ค€๊ณผ ์ธ์ • ํˆฌ์—ฌ ๊ธฐ๊ฐ„์„ ๊ทผ๊ฑฐ ๊ณ ์‹œ์™€ ํ•จ๊ป˜. / ํ”„๋ ˆ๊ฐ€๋ฐœ๋ฆฐ 75 mg ์ฒ˜๋ฐฉ์— ์ƒ๋ณ‘์ด M54.5 ํ•˜๋‚˜์ธ ์ฒญ๊ตฌ ๊ฑด, ๋ˆ„๋ฝ๋œ ์ธ์ •์ƒ๋ณ‘์ด ์žˆ๋Š”์ง€.
  • ์˜๋ฃŒ ๋ฒ•๋ น ๐Ÿ”ง โ€” ์ง„๋ฃŒ๊ธฐ๋ก๋ถ€ ๋ณด์กด๊ธฐ๊ฐ„๊ณผ, ์ „์ž์˜๋ฌด๊ธฐ๋ก์„ ์™ธ๋ถ€ ํด๋ผ์šฐ๋“œ์— ๋ณด๊ด€ํ•  ๋•Œ์˜ ์š”๊ฑด์„ ๊ทผ๊ฑฐ ์กฐ๋ฌธ๊ณผ ํ•จ๊ป˜.

๐Ÿง  Clinical reasoning

  • Differential diagnosis โ€” 45F with lupus nephritis on mycophenolate and prednisone, fever, dry cough, bilateral ground-glass opacities, CD4 180. Differential and workup?
  • ADR causality โ€” 80์„ธ ์—ฌ์„ฑ, moxifloxacin 400 mg IV ํˆฌ์—ฌ ์ค‘ ์ „์‹  ์†Œ์–‘๊ฐ ๋ฐœ์ƒ, ์ค‘๋‹จ ํ›„ ํ˜ธ์ „, ์žฌํˆฌ์—ฌ ์—†์Œ. WHO-UMC ๊ธฐ์ค€์œผ๋กœ ์ธ๊ณผ๊ด€๊ณ„ ํ‰๊ฐ€.
  • Clinical calculation โ€” 72์„ธ ๋‚จ์„ฑ, 68 kg, Cr 1.4 mg/dL. Cockcroft-Gault ๋กœ CrCl ๊ณ„์‚ฐ ํ›„ ์•„ํ”ฝ์‚ฌ๋ฐ˜ ์šฉ๋Ÿ‰ ์กฐ์ • ํ•„์š” ์—ฌ๋ถ€.

๐Ÿ“š Evidence & communication

  • Literature Q&A ๐Ÿ”ง โ€” Does perioperative continuation of SGLT2 inhibitors increase euglycemic DKA risk? Cite every claim.
  • Guideline Q&A ๐Ÿ”ง โ€” 2ํ˜• ๋‹น๋‡จ์—์„œ SGLT2 ์–ต์ œ์ œ๋ฅผ 1์ฐจ ์•ฝ์ œ๋กœ ๊ณ ๋ คํ•  ์ˆ˜ ์žˆ๋Š” ์กฐ๊ฑด์„ ์ง€์นจ ๊ทผ๊ฑฐ์™€ ํ•จ๊ป˜.
  • Patient education โ€” I use insulin daily. How should I store it at home?
  • Clinical documentation โ€” Overnight progress note from labs, vitals, and a stable nursing report.

โš ๏ธ 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 โ€” guidelines, ๊ณ ์‹œ, 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.

๐Ÿ“ 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}
}

๐Ÿค Acknowledgements

This work was supported by the Domain-Specific Foundation Model Project (์ธ๊ณต์ง€๋Šฅ ํŠนํ™” ํŒŒ์šด๋ฐ์ด์…˜ ๋ชจ๋ธ ํ”„๋กœ์ ํŠธ), funded by the Ministry of Science and ICT (๊ณผํ•™๊ธฐ์ˆ ์ •๋ณดํ†ต์‹ ๋ถ€) 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

Consortium Members

๐Ÿ“„ License

This model is licensed under the Apache 2.0 License.

๐Ÿ“ฌ Contact

Downloads last month
314
Safetensors
Model size
30B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for learning-unit/L1-30B-A5B

Finetuned
(3)
this model