GLaDOS model assets
Exact model exports selected for the GLaDOS Rust application. These are transport parts for the built-in model downloader, rather than independently loadable files. The downloader reassembles each asset into the path recorded in models-manifest.json, verifying part and whole-file SHA-256 checksums. Gemma uses five 1 GiB-or-smaller parts. Other assets each use one part.
All 13 selected model roles are included. The GLaDOS voice is distributed under the original publisher’s declared MIT license; see licenses/GLaDOS-voice-NOTICE.txt for the licensing declaration and unresolved third-party provenance.
Contents and profiles
- Local: Gemma-4 E4B Q4_0 and matching Q8_0 multimodal projector, Silero VAD, speech support assets and independent camera assets.
- API: NVIDIA Nemotron English streaming ASR, speech support assets and independent camera assets. No Gemma or Silero download is required for this profile.
- Full: every included role.
- Independent camera: LFM2.5-VL-450M Q4_0 + matching F16 projector, YuNet dynamic detection ONNX and SFace INT8 recognition ONNX. Camera features remain optional in application configuration.
The GLaDOS voice and its metadata are included in both conversation profiles. Runtime libraries ship with the application, not this model bundle. No personal state, settings, credentials, recordings or face identity gallery is included.
Licensing and provenance
See LICENSE, MODEL_SOURCES.json, the embedded per-asset license text in the manifest, and complete notices in licenses/ / model-notices.tar.gz. Licenses differ across assets. The application MIT license does not replace these terms. The GLaDOS voice publisher confirmed MIT licensing intent on 2026-10-22. This records the publisher’s declaration; exact base-checkpoint and third-party training/voice permissions have not been independently established. No Valve or performer permission or endorsement is claimed.
Gemma-4 is Apache-2.0. LFM uses LFM Open License v1.0, which includes commercial-use revenue conditions. Nemotron uses the NVIDIA Open Model License Agreement. The phonemizer uses BSD-3-Clause-Clear and retained source notices. Its modified dictionary is CC-BY-SA-4.0 with attribution and modification notices.
The phonemizer export matches the original GLaDOS release, but that converter did not pin its exact input checkpoint revision. SFace matches OpenCV Zoo's Apache-2.0 licensed INT8 artifact; its complete training/checkpoint lineage remains undisclosed upstream. These are documented provenance limitations, not independent certification of training-data rights.
Versioning
Repository: dnhkng/glados-models. Manifest URLs use the version tag models-2026-10-22. Do not move a published model release tag. Each consumer must receive the trusted SHA-256 of the final manifest independently.
Public, ungated Hugging Face HTTPS downloads are supported by the existing application transport. Private or gated Hugging Face access needs separate token support; the application's GitHub token option is restricted to GitHub and is not a Hugging Face token option.