QAI Hub Models v0.60.0 — Release Notes (2026-08-10)

✨ New Models

Whisper Large v3 Turbo (quantized) (whisper_large_v3_turbo_quantized) — new quantized speech-to-text model
OWL-ViT (owl_vit) — new open-vocabulary object detection model

Gemma-4 E4B and Qwen3-0.6B are now supported via the GenieX QAIRT plugin

Note: Gemma-4 E4B is still under Developer Preview. We're working to further improve on-device quantization and
accuracy. Known issues include repetitive words, formatting glitches, and code-related questions. For more accurate responses,
we recommend the GenieX llama.cpp plugin-based model (quantization technique q4_0).
Qwen3-1.7B — ~15% faster on-device with no accuracy loss; improved response quality on-device
Qwen3-0.6B — improved accuracy on the same-size checkpoint; now ships with [512, 1024, 4096] context-length variants
Fixed Qwen3-8B floating-point model that produced incorrect outputs via evaluate / demo
Fixed regression in LLM export via the qai-hub-models CLI

PT2 export is now the default — most models automatically use torch.export.export
Improved quantized accuracy for:
- PoseNet-MobileNet
- DETR ResNet-50 DC5
- EfficientViT
- UNet Segmentation
- LeViT
- MobileNet v3 Large
Added evaluation support for OpenAI CLIP and DeepBox

New command qai-hub-models install <model> — simpler way to install a model's dependencies
Model name and chipset arguments now support fuzzy matching