The balanced Apple Silicon edition of Miril-DroneVLM-2B-2
Drones can talk.
This repository contains the 8-bit MLX-VLM 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, converted for Apple Silicon. See the primary model card for the caption, answer, location, and pointing responses, fixed system contract, precise-point and coarse-direction rules, 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 Apple unified-memory use while retaining
more weight precision than the 4-bit path. The deployment profile reports
measured artifact size, a recommended memory envelope, and full same-case
quality deltas.
The helper applies the same fixed system prompt and strict response validator as the CUDA path. Invalid or contradictory coordinates are never returned as dispatch-safe points.
Export Integrity
Gemma 4 E2B uses shared key/value layers. The included
shared-KV export audit verifies the expected
tensor-owner layout for this exact MLX artifact; the absence of separate
k_proj, v_proj, and k_norm tensors on sharing layers is intentional.
Complete Benchmark
Deployment comparison
Metric
Merged BF16
MLX 8-bit
Valid JSON
100.0%
100.0%
Schema valid
96.1%
94.6%
Route accuracy
94.8%
93.4%
Caption / answer F1
38.2%
37.3%
Spatial status
79.2%
77.9%
Precise target retained
50.2%
51.2%
Point within 100
33.4%
31.0%
Coarse direction exact
35.0%
32.1%
No-target discipline
93.4%
92.6%
Complete held-out validation
Complete validation comparison
Metric
Merged BF16
MLX 8-bit
Valid JSON
99.9%
100.0%
Schema valid
99.9%
100.0%
Route accuracy
99.9%
100.0%
Caption / answer F1
38.6%
37.8%
Spatial status
86.1%
85.3%
Precise target retained
71.7%
71.2%
Point within 100
38.3%
33.6%
Coarse direction exact
34.4%
33.0%
No-target discipline
96.3%
95.6%
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
MLX 8-bit - Strict cleaned test
Valid JSON
99.8%
100.0%
Schema valid
99.8%
100.0%
Route accuracy
99.8%
100.0%
Reference text F1
60.9%
60.0%
Spatial status
88.0%
86.1%
Coordinate quality
89.7%
88.4%
Precise target retained
65.2%
64.8%
Point within 100
42.6%
39.5%
No-target discipline
96.5%
95.1%
Coverage-matched evidence
Metric
Merged BF16 - Coverage-matched test
MLX 8-bit - Coverage-matched test
Valid JSON
99.9%
100.0%
Schema valid
99.9%
100.0%
Route accuracy
99.9%
100.0%
Reference text F1
64.3%
63.4%
Spatial status
84.2%
82.5%
Coordinate quality
84.9%
83.6%
Precise target retained
63.5%
62.8%
Point within 100
42.8%
39.2%
No-target discipline
96.4%
95.2%
Validation
Strict-cleaned evidence
Metric
Merged BF16 - Strict cleaned validation
MLX 8-bit - Strict cleaned validation
Valid JSON
99.9%
100.0%
Schema valid
99.9%
100.0%
Route accuracy
99.9%
100.0%
Reference text F1
59.9%
59.0%
Spatial status
89.6%
88.5%
Coordinate quality
89.8%
88.9%
Precise target retained
65.8%
65.8%
Point within 100
39.3%
35.0%
No-target discipline
96.7%
96.1%
Coverage-matched evidence
Metric
Merged BF16 - Coverage-matched validation
MLX 8-bit - Coverage-matched validation
Valid JSON
99.9%
100.0%
Schema valid
99.9%
100.0%
Route accuracy
99.9%
100.0%
Reference text F1
62.6%
61.8%
Spatial status
84.5%
83.5%
Coordinate quality
84.5%
83.4%
Precise target retained
60.5%
60.7%
Point within 100
38.4%
33.9%
No-target discipline
96.1%
95.5%
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 MLX and CUDA 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.
