Instructions to use aigensciences/BioGravity-bilingual-Inst with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use aigensciences/BioGravity-bilingual-Inst with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="aigensciences/BioGravity-bilingual-Inst")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("aigensciences/BioGravity-bilingual-Inst")
model = AutoModelForCausalLM.from_pretrained("aigensciences/BioGravity-bilingual-Inst", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use aigensciences/BioGravity-bilingual-Inst with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aigensciences/BioGravity-bilingual-Inst"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aigensciences/BioGravity-bilingual-Inst",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/aigensciences/BioGravity-bilingual-Inst

SGLang

How to use aigensciences/BioGravity-bilingual-Inst with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "aigensciences/BioGravity-bilingual-Inst" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aigensciences/BioGravity-bilingual-Inst",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "aigensciences/BioGravity-bilingual-Inst" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aigensciences/BioGravity-bilingual-Inst",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use aigensciences/BioGravity-bilingual-Inst with Docker Model Runner:
```
docker model run hf.co/aigensciences/BioGravity-bilingual-Inst
```

BioGravity-bilingual-Inst

BioGravity-bilingual-Inst is a Korean–English biomedical instruction model published by AIGEN Sciences, Inc., built on Trillion Labs' Gravity-30B-A5B-0715. It supports biomedical explanations, general instructions, and tool-assisted research through the Biomni A1 interface.

This repository contains the complete model weights, tokenizer, and chat template. Responses use the native <think>...</think> format and explicit A1 action and observation tokens.

Model Summary

Property	Value
Publisher	AIGEN Sciences, Inc.
Upstream checkpoint	`trillionlabs/Gravity-30B-A5B-0715`
Total parameters	29.56B
Active parameters	Approximately 5.34B, counting embedding tables
Architecture	Native Transformers `DeepseekV3ForCausalLM`; MLA and sparse MoE
Transformer layers	52
Routed experts	64 per MoE layer; 8 selected per token
Primary languages	Korean and English
Configured context limit	131,072 tokens
Evaluated serving context	49,152 tokens
Stored precision	FP32; 24 safetensors shards, approximately 118.3 GB
Evaluated inference precision	BF16
Model vocabulary / tokenizer size	151,552 / 151,371
A1 interface	`<execute>`, `<solution>`, and `<observation>`
Verified Transformers version	4.57.6

The configured context limit is an architectural setting. This release was evaluated with a 49,152-token serving window; performance at the full configured limit has not been measured. BF16 loading uses about 59 GB for weights, plus KV cache and runtime memory.

Quickstart

Installation

Use a CUDA-compatible PyTorch installation for your GPU, then install the tested Transformers version and loading dependencies:

pip install "transformers==4.57.6" accelerate safetensors

Using Transformers

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "aigensciences/BioGravity-bilingual-Inst"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
    attn_implementation="sdpa",
).eval()

messages = [
    {"role": "system", "content": "You are a helpful biomedical research assistant."},
    {"role": "user", "content": "관찰연구에서 상관관계와 인과관계의 차이를 한국어로 설명해 주세요."},
]
inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    return_tensors="pt",
    return_dict=True,
    tokenizer_kwargs={"return_token_type_ids": False},
).to(model.device)

with torch.inference_mode():
    outputs = model.generate(
        **inputs,
        max_new_tokens=1024,
        do_sample=False,
        pad_token_id=tokenizer.pad_token_id,
        eos_token_id=tokenizer.eos_token_id,
    )

continuation = outputs[0, inputs["input_ids"].shape[-1]:]
print(tokenizer.decode(continuation, skip_special_tokens=False))

The checkpoint loads through the native Transformers DeepSeek-V3 implementation. Its configuration, expert layout, tokenizer, and RoPE settings form the verified release and should be kept together.

Thinking mode

By default, add_generation_prompt=True leaves the prompt inside an open <think> segment. The generated continuation supplies the remaining text, including </think> and the answer or action.

The included template accepts enable_thinking=False to prefill an empty segment:

inputs = tokenizer.apply_chat_template(
    messages,
    add_generation_prompt=True,
    enable_thinking=False,
    return_tensors="pt",
    return_dict=True,
    tokenizer_kwargs={"return_token_type_ids": False},
).to(model.device)

A1 tool use

Message purpose	Format
Code action	`<think>...</think><execute>Python code</execute>`
Tool observation	`<observation>tool result</observation>`
Final answer	`<think>...</think><solution>answer</solution>`

The harness executes code actions, appends observations, and asks the model for its next decision. It stops action generation at </execute> or </solution>, preserving the stop marker. The model repository supplies the policy; the A1 runtime supplies tools, the execution environment, and the data lake.

Set skip_special_tokens=False when decoding or serving. The six A1 protocol tokens are registered as special tokens; preserving them retains the action boundaries required by the harness.

The inherited template also renders tools= schemas and XML <tool_call> messages. The A1 integration uses the execute/solution protocol.

Deployment

vLLM

Evaluation used vLLM 0.19.0 with one B200 per model replica. The following configuration follows that serving setup. Use a separate environment from training so vLLM can install its matching PyTorch dependencies.

pip install "vllm==0.19.0" "transformers==4.57.6"

VLLM_MLA_DISABLE=1 VLLM_WORKER_MULTIPROC_METHOD=spawn \
python -m vllm.entrypoints.openai.api_server \
  --model aigensciences/BioGravity-bilingual-Inst \
  --served-model-name BioGravity-bilingual-Inst \
  --host 127.0.0.1 --port 8000 \
  --dtype bfloat16 --tensor-parallel-size 1 \
  --max-model-len 49152 \
  --gpu-memory-utilization 0.85 \
  --max-num-seqs 32 --max-num-batched-tokens 4096 \
  --attention-backend TRITON_ATTN \
  --compilation-config '{"mode":0,"cudagraph_mode":"FULL_DECODE_ONLY","cudagraph_capture_sizes":[1,2,4,8,16,24,32],"cudagraph_specialize_lora":false}'

A general chat request:

curl http://127.0.0.1:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "BioGravity-bilingual-Inst",
    "messages": [{"role": "user", "content": "Explain correlation and causation in observational studies."}],
    "temperature": 0,
    "max_tokens": 1024,
    "skip_special_tokens": false
  }'

For A1 requests, the harness additionally supplies "stop": ["</execute>", "</solution>"] and "include_stop_str_in_output": true and handles the action/observation loop.

Acknowledgements

We acknowledge Trillion Labs for the upstream Gravity model and the Biomni project for its biomedical agent and evaluation framework. The organization of this card follows the Gravity-30B-A5B-Preview model card, with model-specific specifications for this release.

Citation

@misc{aigensciences2026biogravitybilingualinst,
  title = {BioGravity-bilingual-Inst},
  author = {{AIGEN Sciences, Inc.}},
  year = {2026},
  url = {https://huggingface.co/aigensciences/BioGravity-bilingual-Inst}
}

Contact

Organization: AIGEN Sciences, Inc.
Model discussions: BioGravity-bilingual-Inst discussions

Downloads last month: 154

Safetensors

Model size

30B params

Tensor type

F32

Model tree for aigensciences/BioGravity-bilingual-Inst

Base model

trillionlabs/Gravity-30B-A5B-0715

Finetuned

(1)

this model