Instructions to use MirilAI/Miril-DroneVLM-2B-2-bnb4 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use MirilAI/Miril-DroneVLM-2B-2-bnb4 with Transformers:

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

pipe = pipeline("image-text-to-text", model="MirilAI/Miril-DroneVLM-2B-2-bnb4")
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)

# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("MirilAI/Miril-DroneVLM-2B-2-bnb4")
model = AutoModelForMultimodalLM.from_pretrained("MirilAI/Miril-DroneVLM-2B-2-bnb4", 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=40)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use MirilAI/Miril-DroneVLM-2B-2-bnb4 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "MirilAI/Miril-DroneVLM-2B-2-bnb4"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "MirilAI/Miril-DroneVLM-2B-2-bnb4",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Use Docker

docker model run hf.co/MirilAI/Miril-DroneVLM-2B-2-bnb4

SGLang

How to use MirilAI/Miril-DroneVLM-2B-2-bnb4 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 "MirilAI/Miril-DroneVLM-2B-2-bnb4" \
    --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": "MirilAI/Miril-DroneVLM-2B-2-bnb4",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

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 "MirilAI/Miril-DroneVLM-2B-2-bnb4" \
        --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": "MirilAI/Miril-DroneVLM-2B-2-bnb4",
		"messages": [
			{
				"role": "user",
				"content": [
					{
						"type": "text",
						"text": "Describe this image in one sentence."
					},
					{
						"type": "image_url",
						"image_url": {
							"url": "https://cdn.britannica.com/61/93061-050-99147DCE/Statue-of-Liberty-Island-New-York-Bay.jpg"
						}
					}
				]
			}
		]
	}'

Docker Model Runner
How to use MirilAI/Miril-DroneVLM-2B-2-bnb4 with Docker Model Runner:
```
docker model run hf.co/MirilAI/Miril-DroneVLM-2B-2-bnb4
```

Miril-DroneVLM-2B-2-bnb4

The compact CUDA edition of Miril-DroneVLM-2B-2

Drones can talk.

This repository contains the 4-bit bitsandbytes/NF4 deployment variant of Miril-DroneVLM-2B-2, Miril.ai's open-weight aerial VLM. Give it an overhead image and an ordinary question such as “What am I looking at?”, “Any people?”, “Choose a place for this parcel,” or “Track the white car.” It returns typed JSON that identifies how the answer should be interpreted.

This is the same four-route model interface as the merged checkpoint, packaged for compact CUDA deployment. Behavior is defined by the primary model card. It covers the caption, answer, location, and pointing responses, exact system contract, coordinate semantics, runnable validation code, WALDO lineage, limitations, and safety requirements.

Methods and citation: Miril-DroneVLM-2B-2: Turning Aerial Detector Labels into a Structured Vision-Language Interface.

Interactive demo: Miril-DroneVLM-2B-2-Demo

Deployment Profile

Artifact size: 6.31 GiB
Recommended starting envelope: 10.5 GiB free VRAM
Estimate covers model weights plus practical single-image runtime headroom; larger images, batching, long generations, and server overhead need more.

Use this variant when CUDA memory matters more than maximum fidelity. The deployment profile reports measured artifact size, a recommended free-VRAM envelope for one-image inference, and full same-case quality deltas against the merged model.

Run It

Download the strict runtime from the primary repository:

hf download MirilAI/Miril-DroneVLM-2B-2 inference.py router_contract.py requirements.txt --local-dir miril-drone-runtime
python -m pip install -r miril-drone-runtime/requirements.txt
python miril-drone-runtime/inference.py \
  --model-id MirilAI/Miril-DroneVLM-2B-2-bnb4 \
  --image drone_frame.jpg \
  --prompt "Track the white car."

The exported checkpoint carries its quantization configuration. Do not add --load-4bit when loading this already-quantized repository.

Use transformers>=5.12.1. This artifact retains Gemma 4 E2B's shared-KV layout: language layers 15 through 34 reuse key/value states and intentionally have no separate k_proj, v_proj, or k_norm tensors. The artifact audit records the expected and observed tensor owners.

The helper rejects invalid JSON, schema violations, contradictory status/coordinate combinations, and non-target responses with coordinates. Only a valid target_found + precise_point may become a precise marker. A coarse_grid_direction remains a broad location cue and is never a landing, delivery, pointing, or tracking target.

Complete Benchmark

Metric	Merged BF16	CUDA bnb4
Valid JSON	100.0%	96.7%
Schema valid	96.1%	92.1%
Route accuracy	94.8%	91.2%
Caption / answer F1	38.2%	37.2%
Spatial status	79.2%	69.7%
Precise target retained	50.2%	53.1%
Point within 100	33.4%	29.7%
Coarse direction exact	35.0%	30.2%
No-target discipline	93.4%	87.5%

Complete held-out validation

Metric	Merged BF16	CUDA bnb4
Valid JSON	99.9%	99.0%
Schema valid	99.9%	98.7%
Route accuracy	99.9%	98.6%
Caption / answer F1	38.6%	37.2%
Spatial status	86.1%	79.8%
Precise target retained	71.7%	73.6%
Point within 100	38.3%	34.0%
Coarse direction exact	34.4%	31.2%
No-target discipline	96.3%	92.4%

Cleaned held-out deployment audit

After training, a stricter held-out audit removed pointing rows whose targets fall below the model-visible size threshold, then ran every release artifact on the complete revised validation and test splits. Strict-cleaned rows use only accepted evidence. Coverage-matched rows add evidence-preserving questions on the same held-out images to restore the earlier route and pointing action/status mix; they do not recreate the earlier object-class histogram.

Final held-out test

Strict-cleaned evidence

Metric	Merged BF16 - Strict cleaned test	CUDA bnb4 - Strict cleaned test
Valid JSON	99.8%	89.3%
Schema valid	99.8%	89.2%
Route accuracy	99.8%	89.1%
Reference text F1	60.9%	53.5%
Spatial status	88.0%	72.0%
Coordinate quality	89.7%	78.9%
Precise target retained	65.2%	60.3%
Point within 100	42.6%	34.3%
No-target discipline	96.5%	85.1%

Coverage-matched evidence

Metric	Merged BF16 - Coverage-matched test	CUDA bnb4 - Coverage-matched test
Valid JSON	99.9%	89.4%
Schema valid	99.9%	89.3%
Route accuracy	99.9%	89.1%
Reference text F1	64.3%	56.6%
Spatial status	84.2%	70.8%
Coordinate quality	84.9%	74.7%
Precise target retained	63.5%	60.0%
Point within 100	42.8%	34.4%
No-target discipline	96.4%	84.1%

Validation

Strict-cleaned evidence

Metric	Merged BF16 - Strict cleaned validation	CUDA bnb4 - Strict cleaned validation
Valid JSON	99.9%	96.4%
Schema valid	99.9%	96.1%
Route accuracy	99.9%	96.0%
Reference text F1	59.9%	56.8%
Spatial status	89.6%	80.2%
Coordinate quality	89.8%	85.5%
Precise target retained	65.8%	67.8%
Point within 100	39.3%	34.6%
No-target discipline	96.7%	91.8%

Coverage-matched evidence

Metric	Merged BF16 - Coverage-matched validation	CUDA bnb4 - Coverage-matched validation
Valid JSON	99.9%	96.3%
Schema valid	99.9%	96.0%
Route accuracy	99.9%	95.9%
Reference text F1	62.6%	59.4%
Spatial status	84.5%	76.6%
Coordinate quality	84.5%	80.1%
Precise target retained	60.5%	62.1%
Point within 100	38.4%	34.0%
No-target discipline	96.1%	90.0%

Validation supports comparison and model selection; test is the final held-out report. These automated scores measure contract and reference agreement, not flight safety.

Spoken-query results on the primary model card apply to the merged BF16 checkpoint. This deployment variant has not inherited that claim without a separate matched audio evaluation.

The merged and deployment variants use identical cases within each evaluation. Automated scores are regression signals, not safety certification.

The tables below compare merged and quantized artifacts on identical complete held-out validation and test cases. Partial runs are excluded. Automated scores are regression signals, not safety certification.

Limits And Safety

This is a research perception model, not a flight controller or certified safety system. Follow the complete limitations and operational guidance on the primary model card.