Miril-Drone-2B-1-bnb4

Miril-Drone-2B-1-bnb4

4-bit CUDA deployment variant of Miril-Drone-2B-1

Drones can talk, and this is the compact CUDA path.

This repository packages the 4-bit bitsandbytes/NF4 variant of Miril-Drone-2B-1, a 2B-class aerial VLM for drone-view imagery.

Use the primary model card for behavior, prompting, schemas, examples, WALDO vocabulary, limitations, and safety notes:

https://huggingface.co/MirilAI/Miril-Drone-2B-1

The V1 prompt contract is the same as the main model: caption_v1, simple_answer_v1, and operational_coordinate_v2. V1 operational coordinates are rough representative grid cues for review, not flight-control commands. Miril-DroneVLM-2B-2 is the newer plain-English typed-router generation.

Interactive demo:

https://huggingface.co/spaces/MirilAI/mirilai-miril-drone-2b-1

Use

Use transformers>=5.12.1. This model retains Gemma 4 E2B's declared shared-KV layout; language layers 15 through 34 intentionally reuse key/value states and do not carry separate k_proj, v_proj, or k_norm tensors.

python eval_generate.py \
  --model_id MirilAI/Miril-Drone-2B-1-bnb4 \
  --processor_id MirilAI/Miril-Drone-2B-1-bnb4 \
  --jsonl eval.jsonl \
  --image_root images \
  --out_jsonl predictions.jsonl \
  --batch_size 1 \
  --max_new_tokens 256

Deployment Profile

<!-- MIRIL_VARIANT_PROFILE_START -->

MIRIL_VARIANT_PROFILE_PENDING

<!-- MIRIL_VARIANT_PROFILE_END -->

Complete Benchmark

<!-- MIRIL_RELEASE_METRICS_START -->

MIRIL_RELEASE_METRICS_PENDING

<!-- MIRIL_RELEASE_METRICS_END -->

The merged and quantized variants are evaluated on identical held-out cases. Automated scores are regression signals, not safety certification.

License

Apache License 2.0. See LICENSE and NOTICE.