This repository contains the 8-bit bitsandbytes 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 lower CUDA memory use. 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.
Estimate covers model weights plus practical single-image runtime headroom; larger images, batching, long generations, and server overhead need more.
Use this variant when you want lower CUDA memory use. Bit width does not imply
that this artifact is behaviorally closer to merged BF16 than the NF4 variant;
the complete same-case tables below are the release evidence. The deployment
profile also reports measured artifact size and a recommended free-VRAM envelope for
one-image inference.
The complete held-out audit shows a clear capability split: this artifact keeps
high JSON, schema, and route reliability, but loses substantial caption,
factual-answer, and coordinate-grounding quality relative to merged BF16. It is
therefore an evaluated compact option, not the recommended high-fidelity CUDA
checkpoint. Use the tables below and strict downstream validation; compare the
NF4 package separately instead of assuming 8-bit is better.
Run It
bash
1hf download MirilAI/Miril-DroneVLM-2B-2 inference.py router_contract.py requirements.txt --local-dir miril-drone-runtime
2python -m pip install -r miril-drone-runtime/requirements.txt
3python miril-drone-runtime/inference.py \4 --model-id MirilAI/Miril-DroneVLM-2B-2-bnb8 \5 --image drone_frame.jpg \6 --prompt "Choose a place to lower this parcel."
The exported checkpoint carries its quantization configuration. The shared helper validates the model's bare JSON before any response is dispatched or drawn.
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.
Complete Benchmark
Deployment comparison
Metric
Merged BF16
CUDA bnb8
Valid JSON
100.0%
99.1%
Schema valid
96.1%
93.6%
Route accuracy
94.8%
91.6%
Caption / answer F1
38.2%
19.6%
Spatial status
79.2%
63.4%
Precise target retained
50.2%
42.4%
Point within 100
33.4%
8.0%
Coarse direction exact
35.0%
9.3%
No-target discipline
93.4%
89.6%
Complete held-out validation
Complete validation comparison
Metric
Merged BF16
CUDA bnb8
Valid JSON
99.9%
99.7%
Schema valid
99.9%
99.7%
Route accuracy
99.9%
99.6%
Caption / answer F1
38.6%
23.0%
Spatial status
86.1%
70.1%
Precise target retained
71.7%
60.7%
Point within 100
38.3%
9.2%
Coarse direction exact
34.4%
11.3%
No-target discipline
96.3%
93.3%
Cleaned held-out deployment audit
Cleaned test comparison
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 bnb8 - Strict cleaned test
Valid JSON
99.8%
99.6%
Schema valid
99.8%
99.6%
Route accuracy
99.8%
99.5%
Reference text F1
60.9%
47.4%
Spatial status
88.0%
76.7%
Coordinate quality
89.7%
84.0%
Precise target retained
65.2%
52.6%
Point within 100
42.6%
9.4%
No-target discipline
96.5%
97.1%
Coverage-matched evidence
Metric
Merged BF16 - Coverage-matched test
CUDA bnb8 - Coverage-matched test
Valid JSON
99.9%
99.6%
Schema valid
99.9%
99.5%
Route accuracy
99.9%
99.5%
Reference text F1
64.3%
50.3%
Spatial status
84.2%
64.1%
Coordinate quality
84.9%
77.6%
Precise target retained
63.5%
49.9%
Point within 100
42.8%
7.8%
No-target discipline
96.4%
97.6%
Validation
Strict-cleaned evidence
Metric
Merged BF16 - Strict cleaned validation
CUDA bnb8 - Strict cleaned validation
Valid JSON
99.9%
99.6%
Schema valid
99.9%
99.5%
Route accuracy
99.9%
99.5%
Reference text F1
59.9%
48.1%
Spatial status
89.6%
80.1%
Coordinate quality
89.8%
84.9%
Precise target retained
65.8%
55.7%
Point within 100
39.3%
9.6%
No-target discipline
96.7%
97.2%
Coverage-matched evidence
Metric
Merged BF16 - Coverage-matched validation
CUDA bnb8 - Coverage-matched validation
Valid JSON
99.9%
99.6%
Schema valid
99.9%
99.5%
Route accuracy
99.9%
99.4%
Reference text F1
62.6%
50.3%
Spatial status
84.5%
66.7%
Coordinate quality
84.5%
78.7%
Precise target retained
60.5%
49.7%
Point within 100
38.4%
8.1%
No-target discipline
96.1%
97.4%
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.
