Skip to content

v0.57.0

Choose a tag to compare

View all tags
@kory kory released this 01 Jul 22:08
v0.57.0
6fd045a
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

🛠️ Introducing the AI Hub Models CLI

This release introduces a new qai-hub-models command-line interface for exploring and using AI Hub Models directly from your terminal. It's a lightweight, standalone tool — a few megabytes with only a handful of small pure-Python dependencies — and most commands run without needing the full qai_hub_models package installed.

  • Discover — models, devices, chipsets, and runtimes list what's available; filter models by quantization, runtime, chipset, and tags.
  • Inspect — info shows a full model overview, perf and numerics show per-device metrics.
  • Fetch — fetch downloads a pre-compiled model asset by chipset/device, runtime, and precision. find searches past releases for assets that match your filters (e.g. QAIRT version).
  • Export & evaluate — export and profile/evaluate models from source using AI Hub Workbench.
pip install qai_hub_models_cli

qai-hub-models models              # list available models
qai-hub-models info mobilenet_v2   # inspect a model
qai-hub-models fetch mobilenet_v2 -r tflite -p float  # download an asset

Run qai-hub-models --help to see all commands.

See the README for a detailed overview: https://github.com/qualcomm/ai-hub-models/blob/v0.57.0/cli/README.md

🚀 Introducing GenieX runtime for LLMs and VLMs

This release adds support for GenieX, an on-device generative-AI runtime for running frontier LLMs and VLMs locally on Qualcomm devices across the NPU, GPU, and CPU. GenieX runs on Windows ARM64, Android, and Linux.

GenieX is now available in AI Hub Models through two new target runtimes:

  • geniex_qairt — QAIRT-backed and NPU-only, compiled with AI Hub Workbench for the highest on-device performance.
  • geniex_llamacpp — llama.cpp-backed, runs any GGUF model straight from Hugging Face across NPU/GPU/CPU (q4_0 recommended for best Hexagon NPU support).

Performance numbers for both GenieX backends are now published across our LLM lineup, collected via the scorecard on Windows, Linux, and Android.

Learn more: https://github.com/qualcomm/geniex

Changelog

New Runtimes

  • Two new TargetRuntime values added for GenieX:
    • GENIEX_QAIRT — GenieX LLM runtime backed by QAIRT.
    • GENIEX_LLAMACPP — GenieX LLM runtime backed by llama.cpp / GGUF; does not compile through AI Hub Workbench and introduces new GGUF precisions mxfp4, q8_0, q4_0.

New Models

  • VLMs & LLMs (GenieX accelerated by QAIRT)
    • Qwen3-8B
    • Qwen3-VL-4B-Instruct
  • VLMs & LLMs (GenieX accelerated by Llama CPP)
    • Qwen3-0.6B
    • Qwen3-1.7B
    • Qwen3.5-0.8B
    • Qwen3.5-2B
    • Qwen3-VL-2B-Instruct
    • Qwen3-VL-4B-Instruct
    • Gemma-4-E2B-IT
    • Gemma-4-E4B-IT
    • GPT-OSS-20B
    • Granite-4.0-Micro
    • Ministral-3.3B-Instruct-2512
    • Phi-4-Mini-Instruct
  • Vision
    • EfficientNet-Lite4
    • NAFSSR
    • SAM3
    • YOLO26-Pose
  • Embedding Models
    • MiniLM-v2

New CLI

  • The AI Hub Models CLI (qai-hub-models --help) can now search the model / device catalog and display all model information that is available on our website.
  • New syntax to run export / evaluate locally: qai_hub_models export <model> <args>

Deprecations

  • Running export / evaluate scripts (python -m qai_hub_models.models.<id>.export) directly has been deprecated and will be removed in a future release. Use the CLI instead: qai_hub_models export <model>
  • The Genie runtime has been deprecated (use GenieX instead!). Pre-exported Genie assets and performance / accuracy data will not be updated with future releases of AI Hub Models (the existing downloads and data will continue to be available in releases during the deprecation period). Support will be fully removed in a future release.

Other Improvements

  • LLM modernization — Qwen3 models (4B, 8B, 4B-Instruct-2507) and the Llama family migrated from the legacy static deployment to the dynamic-shape (dynamo) pattern. Dynamic shapes let a single export cover all context/sequence lengths instead of recompiling per shape, significantly reducing export times.
  • New set14 dataset for super-resolution; RangeNet++ and FastSAM evaluation support added.
  • NPU support added for CREStereo, EyeGaze, and MoveNet (not yet reflected in this release's performance/accuracy data).

Bug / accuracy fixes

  • RTMDet & YOLOv6 accuracy gaps — added a new COCO detection evaluator and corrected model post-processing so mAP now matches published baselines (YOLOv6 ~37.4 vs 37.5 published; RTMDet ~48.7 vs 49.4 published on 5000 samples).
  • w8a8 enablement for ResNet-Mixed, Sequencer2D, NAFNet-Deblur, and ConvNeXt-Tiny — each validated to quantize with < 2-point accuracy drop vs. the torch baseline (on-device, 300 samples).
  • Fixed wikitext perplexity for Qwen3-4B and Qwen3-8B.

Performance Numbers

  • Added GenieX performance data (Llama.cpp and QAIRT) across the LLM lineup.
Assets 2
Loading