GLM-5.3 is live. Run Z.ai's latest model on Friendli Model APIs. Try it today ➜
SNUH-HARI
Available on FriendliAI
Run this model inference on single tenant GPU with unmatched speed and reliability at scale.
Model Details
Model Provider
SNUH-HARI
Model Tree
snuh/hari-q3-8b
Input Modalities
Output Modalities
Supported Functionality
| hari-q3-8b | + this adapter | Qwen3.8 27B | Qwen3.8-Flash-Next 180B-A6B | |
|---|---|---|---|---|
| exact % | 46.1 | 93.1 | 90.7 | 92.7 |
| unsafe execution % (82 reject/clarify items) | 43.9 | 0.0 | 9.8 | 15.9 |
| stop recall % (16 items) | 100 | 100 | 100 | 100 |
| over-refusal % (413 execute items) | 0.2 | 0.2 | 4.1 | 0.2 |
| median latency | 0.52 s | 0.62 s | 2.71 s (API) | 0.68 s (API) |
vitaldb/qwen3-asr-1.7b-kor-or-lora): 66.9%.python
from vllm import LLM, SamplingParamsfrom vllm.lora.request import LoRARequestfrom vllm.sampling_params import StructuredOutputsParams llm = LLM("snuh/hari-q3-8b", enable_lora=True, max_lora_rank=64, max_model_len=4096)tok = llm.get_tokenizer()messages = [{"role": "system", "content": scene_prompt}, # see code/bench/prompt.py in the dataset repo {"role": "user", "content": '[직전 로봇 동작] 없음\n[집도의 발화] "IMV, 아 아니다, IMA 쪽 보여줘"'}]prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)sp = SamplingParams(temperature=0, max_tokens=400, structured_outputs=StructuredOutputsParams(json=schema))out = llm.generate([prompt], sp, lora_request=LoRARequest("cmd", 1, "SNUH-HARI/hari-q3-8b-surgical-command-lora"))The system prompt lists the scene's procedure, phase, available actions, target names and interpretation rules; the
same prompt was used for training and evaluation (code/bench/prompt.py, schema in code/bench/schema.py).
train/train_v0.2.jsonl).This work was supported by the Korea ARPA-H Project through the Korea Health Industry Development Institute (KHIDI), funded by the Ministry of Health and Welfare, Republic of Korea (RS-2025-25397510).
from vllm import LLM, SamplingParamsfrom vllm.lora.request import LoRARequestfrom vllm.sampling_params import StructuredOutputsParams llm = LLM("snuh/hari-q3-8b", enable_lora=True, max_lora_rank=64, max_model_len=4096)tok = llm.get_tokenizer()messages = [{"role": "system", "content": scene_prompt}, # see code/bench/prompt.py in the dataset repo {"role": "user", "content": '[직전 로봇 동작] 없음\n[집도의 발화] "IMV, 아 아니다, IMA 쪽 보여줘"'}]prompt = tok.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False)sp = SamplingParams(temperature=0, max_tokens=400, structured_outputs=StructuredOutputsParams(json=schema))out = llm.generate([prompt], sp, lora_request=LoRARequest("cmd", 1, "SNUH-HARI/hari-q3-8b-surgical-command-lora"))