Instructions to use MeTask-Rain/metask-jev-rain-12B-trained with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MeTask-Rain/metask-jev-rain-12B-trained with Transformers:

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

pipe = pipeline("text-generation", model="MeTask-Rain/metask-jev-rain-12B-trained")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("MeTask-Rain/metask-jev-rain-12B-trained")
model = AutoModelForMultimodalLM.from_pretrained("MeTask-Rain/metask-jev-rain-12B-trained", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.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=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use MeTask-Rain/metask-jev-rain-12B-trained with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MeTask-Rain/metask-jev-rain-12B-trained"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MeTask-Rain/metask-jev-rain-12B-trained",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/MeTask-Rain/metask-jev-rain-12B-trained

SGLang

How to use MeTask-Rain/metask-jev-rain-12B-trained 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 "MeTask-Rain/metask-jev-rain-12B-trained" \
    --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": "MeTask-Rain/metask-jev-rain-12B-trained",
		"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 "MeTask-Rain/metask-jev-rain-12B-trained" \
        --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": "MeTask-Rain/metask-jev-rain-12B-trained",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use MeTask-Rain/metask-jev-rain-12B-trained with Docker Model Runner:
```
docker model run hf.co/MeTask-Rain/metask-jev-rain-12B-trained
```

metask-jev-rain-12B-trained

MeTask's 12B model makes structured decisions: given a state and a question, the service returns a judgment and probabilities for the available options. Our trained adapter, together with the base checkpoint it runs on, is available for download and private deployment. On JevBench, it scores 77.9 and ranks #6 in the qualifying open-weight capability ranking as of 2026-10-17; it is the highest-ranked entry labeled 12B in the system name.

Organizer-measured results

Results below are organizer-measured Capability Scores, checked on 2026-10-17 against JevBench board v1.7.46 (page updated 2026-10-16). Ranks refer to the 129 open-weight systems qualifying for the capability ranking under the official cost and latency limits, not the composite-score ranking or every listed system. The leaderboard's 'official #' label refers to its separate composite ranking.

For these full-set results, the operator measures each system on 1,500 decisions: 1,200 sealed items that submitters cannot access and 300 public items.

Evaluated service	Trained	Capability Score	Capability rank	Checked
metask-jev-rain-12B-trained	yes	77.9	6; highest among entries labeled 12B in the system name	2026-10-17

The organizer reports a 0.115 s on-GPU p50 for this service, excluding network transit; its standard self-hosted adjustment gives 0.38 s (2 × measured p50 + 0.15 s). These are the leaderboard's measurements and conversion, not our local A100 measurements. See the linked leaderboard for the measurement basis.

Data and evaluation

The adapter was trained on our own data, not on the 231 public JevBench items. The 231 public items available to us at the time were used to fit service routing and calibration parameters and for service self-tests and regression/reproducibility checks; the organizer's score is measured on 1,200 sealed items plus 300 public items. These public-item checks are in-sample service checks, not independent evidence of generalization or the official benchmark result.

See the service disclosure for data use and measurement conditions. The organizer-measured results above evaluate the complete service; a different runtime or calibration setup may give different results with the same weights.

Use and deployment

We trained this model for decision tasks on Google's open-weight Gemma 4 12B base. The download contains the Gemma 4 base checkpoint and our trained decision adapter; both must be loaded together, and loading the base alone omits our training.

Use the service package to reproduce the structured decision interface. Both the base checkpoint and adapter/ are required. The service uses vLLM 0.31.0 and applies the adapter on every request. It may use a short second reasoning pass for some decisions; it is not always a single-pass system.

HF-generated Transformers examples may load only the base checkpoint; follow the service instructions to include our adapter and structured decision interface.

Base revisions, adapter structure, quantization and the serving example are in Technical notes. Evaluate decision accuracy and probability calibration on your own tasks before deployment.

License and attribution

Apache-2.0, following the base model's Gemma 4 license. Base-model attribution is retained in this card's metadata and the technical notes. Retain the applicable license and attribution notices when redistributing.

Contact: chenyuwang166@gmail.com