Qualcomm AI Hub Models
QualcommQualcomm's catalog of machine learning models optimized for Snapdragon and other Qualcomm devices, browsable at aihub.qualcomm.com and installable as the qai-hub-models package. A CLI and Python API browse the catalog and fetch deployable assets per runtime and precision, and demos run locally in PyTorch. Compiling and profiling on real devices goes through the hosted Qualcomm AI Hub Workbench.
Qualcomm AI Hub Workbench, the hosted compile and profiling service, has no entry of its own; the features that depend on it are why the openness reading is open core. The package metadata still links quic/ai-hub-models, an earlier name of the same repository.
Openness
4 medium confidence- license
- BSD-3-Clause(OSI)
- source
- public(github.com/qualcomm/ai-hub-models)
- core features withheld
- yes — compilation, on-device profiling and cloud-device inference need the closed AI Hub Workbench and an API token
The catalog code is BSD-3-Clause and its demos run locally, but compilation, on-device profiling and cloud-device inference go through Qualcomm's closed AI Hub Workbench with an API token. Workbench is free today, yet those features cannot be built from the published source.
- https://raw.githubusercontent.com/qualcomm/ai-hub-models/HEAD/LICENSE recorded 2026-09-27
BSD 3-Clause text, "Copyright 2025 Qualcomm Technologies, Inc. and/or its subsidiaries."
- https://raw.githubusercontent.com/qualcomm/ai-hub-models/HEAD/README.md recorded 2026-09-27
"Many features of AI Hub Models (such as model compilation, on-device profiling, etc.) require access to Qualcomm® AI Hub Workbench"; "All end-to-end demos can also run locally via PyTorch".
- https://workbench.aihub.qualcomm.com/docs/hub/faq.html recorded 2026-09-27
"Qualcomm® AI Hub Workbench, our platform, is currently completely free to use."
Adoption
1 medium confidenceInstalls of the qai-hub-models package from PyPI are the catalog's own download channel, and the last month is well below the spring, when monthly installs ran several times higher. The figure understates use of the catalog, since the web page and Qualcomm's Hugging Face organization serve assets that never pass through the package.
- https://huggingface.co/api/organizations/qualcomm/overview recorded 2026-09-27
"fullname":"Qualcomm", "isVerified":true, "numModels":257 on the Hugging Face organization.
- https://packages.ecosyste.ms/api/v1/registries/pypi.org/packages/qai-hub-models recorded 2026-09-27
"downloads":31663 over "last-month", from a record last refreshed 2026-07-15 that still lists release 0.57.3.
- https://pypistats.org/api/packages/qai-hub-models/overall?mirrors=false recorded 2026-09-27
Daily downloads without mirrors; summed by month they run about 34,000 to 36,000 in April to June 2026, 24,900 in July and 16,952 in August.
- https://pypistats.org/api/packages/qai-hub-models/recent recorded 2026-09-27
"last_month":5231, "last_week":1300 for qai-hub-models, 2026-09-27.
Capability
2 high confidenceLike the NGC catalog, it hosts ready-to-run assets and serves them by name, with a CLI and API to browse and fetch them. Only Qualcomm publishes, and new models need internal legal approval, so it is an operator's own collection rather than an open registry.
- https://aihub.qualcomm.com/models recorded 2026-09-27
"All Models Optimized and Validated by Qualcomm"; "541 model variants (235 models)".
- https://raw.githubusercontent.com/qualcomm/ai-hub-models/HEAD/CONTRIBUTING.md recorded 2026-09-27
"Legal approval is required before publishing new models. Submit a request at go/genairequest."
- https://raw.githubusercontent.com/qualcomm/ai-hub-models/HEAD/README.md recorded 2026-09-27
"qai-hub-models models # browse the catalog" and "qai-hub-models fetch mobilenet_v2 --runtime tflite --precision float # download a deployable asset"; "The CLI also offers a Python API."
Verified 2026-09-27