Deployment-ready mobile vision-language encoder assets for the Memoria project, including ONNX/TFLite/NCNN exports, validation scripts, and release tooling.

This release focuses on Memoria's mobile deployment scenario. It does not claim to be the first public ONNX/TFLite export of MobileCLIP or MobileCLIP2.

Released Models

Model	ONNX	TFLite	NCNN	Text Encoder	Validation
mobileclip-s2	137.6 MB	68.4 MB float16	6 files, 378.2 MB	-	ONNX cosine >= 0.99999988; NCNN cosine >= 0.99999988 vs ONNX
mobileclip2-s0	43.4 MB	-	-	242.3 MB ONNX	ONNX cosine >= 0.99999988
mobileclip2-s2	136.4 MB	-	-	242.3 MB ONNX	ONNX cosine >= 0.99999994
mobileclip2-b	329.5 MB	-	-	242.5 MB ONNX	ONNX cosine >= 0.99999988
mobileviclip-small-frame-encoder	67.3 MB	-	raw + optimized NCNN, 67.5 MB	-	ONNX-vs-NCNN cosine >= 0.9997077584

Validation

Model	Backend	Min Cosine	Max Abs Error	Status
mobileclip-s2	ONNX	0.99999988	8.05e-07	passed
mobileclip-s2	TFLite fp16	0.99997151	1.38e-03	passed
mobileclip-s2	NCNN image encoder	0.99999988	1.47e-04	passed
mobileclip2-s0	ONNX	0.99999988	2.98e-07	passed
mobileclip2-s2	ONNX	0.99999994	5.36e-07	passed
mobileclip2-b	ONNX	0.99999988	2.15e-06	passed
mobileviclip-small-frame-encoder	ONNX-vs-NCNN frame feature-map	0.9997077584	1.745875776e-01	passed

ONNX and available TFLite artifacts were numerically validated against the PyTorch reference model. NCNN artifacts were validated against their ONNX exports.

Reproducibility

Each model directory contains:

model.onnx
model.int8.tflite, when supported and present
model.ncnn.param / model.ncnn.bin, when supported and present
config.json
metrics.json
README.md
scripts/ with export, validation, quantization, and upload helpers

Build v0.1 locally from the Memoria repository:

python ai_tools/mobile_vision_release/export_mobile_vision_assets.py --manifest ai_tools/mobile_vision_release/release_manifest.example.json --run-export --run-validate

Upload only after reviewing metrics.json and the generated README files:

python ai_tools/mobile_vision_release/upload_to_hf.py --release-root build/mobile_vision_release --repo-id <owner>/<repo> --create-repo --execute

MobileViCLIP v0.3 Status

mobileviclip-small-frame-encoder-ncnn is a verified spatial/frame feature-map NCNN subgraph for MobileViCLIP-small. It is not a full MobileViCLIP-small video encoder.

Release tag: v0.3-mobileviclip-small-frame-encoder-ncnn

The full-video pnnx NCNN graph is blocked and not released: it loads successfully, but crashes at extract(out0) in both Python NCNN and native Android C++ NCNN runtimes. The v0.3 artifact includes only the first-stage frame feature extractor. Temporal head, video aggregation, projection, and L2 normalization are expected to run in Python/NumPy/Android post-processing or in a future separately validated subgraph.

Original repo: https://github.com/MCG-NJU/MobileViCLIP

Team ONNX reference: https://huggingface.co/szw0407/mobileviclip_small_onnx

Attribution

Exported by: Memoria project
Quantized by: Memoria project
Original model licenses and source links are listed in each model directory.
Current model artifacts are derived from Apple MobileCLIP / MobileCLIP2 and use the upstream Apple Sample Code License (apple-amlr) metadata.

No Hugging Face tokens, API keys, local absolute paths, or private data should be committed or uploaded with these release folders.

youthfedpycharm/memoria-mobile-vision-assets

Memoria Mobile Vision Assets

Released Models

Validation

Reproducibility

MobileViCLIP v0.3 Status

Attribution

youthfedpycharm/
memoria-mobile-vision-assets