v0.61.0

New Models

• WeDetect (wedetect) - text-conditioned open-vocabulary object detection.
• SigLIP2 (siglip2) - image/text embedding model.
• RepViT (repvit) - pure-CNN ImageNet classifier with ViT-inspired design.
• FunASR-Conformer-EN (funasr_conformer_en) - English speech recognition.
• OWL-V2 (owlv2) - open-vocabulary object detection.
• MobileFaceNet (mobile_facenet) - lightweight real-time face verification.
• OSNet (osnet) - person re-identification embeddings for matching identities across camera views.

LLMs

• New Grace2 accuracy benchmark. This is a benchmark that we developed to better evaluate the
  post-quantization on-device accuracy of LLMs. Prompts of various categories are judged by
  a larger LLM. See grace.md.
• Quantization recipes are now published in each LLM's manifest.yaml - precision layout (blocks, lm_head, KV cache, activations)
  and recipe steps (SpinQuant, AdaScale, calibration datasets).
• Added QCS8275 assets across most LLMs, and SA8295P for Llama 3.2 1B/3B (w4).
• Fixed Gemma-4 RoPE tensor names and a Genie PLE LUT >4 GiB failure.
• Fixed the chat template being dropped from Genie/GenieX bundles with recent transformers versions.
• Fixed Qwen3-8B MMLU evaluation running out of memory on 80 GiB GPUs.

Model improvements & fixes

• Added w8a16 to ShuffleNet-v2.
• EfficientNet-B0 and MobileNet-v3-Small moved from w8a16 to w8a16_mixed_int16 to fix on-target accuracy.
• YOLOv11-Pose: w8a16 is now the only quantized precision.
• YOLOE-Seg and YOLOv8-OBB: default quantized precision is now w8a16 instead of w8a8.
• TrOCR switched to the trocr-small-handwritten checkpoint.
• Fixed a Detectron2-Detection regression (KITTI dataset issue).
• PointNet exports now bundle modelnet10_labels.txt.
• Fixed CenterPoint import failure on CPU-only hosts.

CLI

• qai-hub-models install <model> replaces pip install "qai-hub-models[<model>]" in all model READMEs.
• Fixed installs leaving both onnxruntime and onnxruntime-gpu present.
• AIMET-ONNX now installs from PyPI, minimum version 2.28 (which extends support to Python 3.10+).

Other

• Downloadable assets are now restricted for FastSAM-S, FastSAM-X, UNet Segmentation, per their licenses.

Performance numbers

• Added VENTUNO Q performance numbers for LLMs and VLMs (24 models).
• Performance and accuracy data refreshed across the catalog.