NOTICE — upstream attributions

Base model
----------
VisionPsy-Nano-460M and VisionPsy-Nano-460M-Flash by QVAC.
  HF:    https://huggingface.co/qvac/VisionPsy-Nano-460M
  Code:  https://github.com/tether-ai-research/qvac-visionpsy-nano
  Blog:  https://huggingface.co/blog/qvac/visionpsy

All MY_* GGUF artifacts in this repository's GitHub Releases are quantized
derivatives of QVAC's model weights (domain-calibrated imatrix quantization
and/or embedding-protected quantization). They inherit QVAC's license terms;
consult the upstream repository before redistribution. QVAC's own artifact
names appear in tables and charts only as within-harness references.

Runtime
-------
llama.cpp with QVAC's patched mtmd fork (multimodal batch inference).
  https://github.com/ggml-org/llama.cpp

Evaluation
----------
VLMEvalKit (strict judge-free scoring), MMStar, ScienceQA_TEST, OCRBench,
PerceptionBench (MoonshotAI), MMMU, MMBench.
  https://github.com/open-compass/VLMEvalKit
  https://github.com/MoonshotAI/PerceptionBench

Judges (via OpenRouter): qwen3.6-27b (QVAC's published judge protocol),
gpt-oss-120b (PerceptionBench scoring).