Model testing
Choose an internal hospital test or a hosted comparison of self-hostable models.
Hospital models · internal only
Run Halo inside the hospital network, then connect your installed transcription and language models. Try synthetic text or audio; results stay separate from your draft and style history.
Hosted comparison · DeepInfra + OpenAI
One recording, five report results. Audio is sent to DeepInfra and OpenAI. Your normal report, selected AI engine and learned style are not changed.
Models and connection
- Qwen3-ASR-1.7B × Qwen3.8-27B
- Qwen3-ASR-1.7B × gpt-oss-20b
- Whisper large-v3 × Qwen3.8-27B
- Whisper large-v3 × gpt-oss-20b
- gpt-4o-mini-transcribe × GPT-5.6 Sol (API baseline)
On the live website, the administrator installs the provider key in a server secret file. No hospital endpoint is required. ChatGPT web is a separate product.
Two open STT models are each run once and reused across the four combinations. The OpenAI baseline uses its own STT. The same style snapshot and Conclusion option are used throughout. Models use a fixed 8,192-token output cap; open models use low reasoning, Sol uses none. Provider versions/precision may differ from a future hospital deployment.
Raw transcripts
Review medical terminology, negation, laterality, measurements, omissions and invented findings. Cloud timing is not hospital GPU timing. Temporary server results expire after 30 minutes; download them before leaving if needed.
Report
English and Korean · Plain textLoading Sprint library…
Request Halo AI
Ask for a revision to the current report, or describe a style to change the current report. Your feedback will help guide your AI style. Revisions support Undo. Keep Remember style preferences checked for reusable style requests; uncheck it for a one-off change. Example: “Use 2026-12-25 instead of December 25th, 2026 for dates.”
A second pair of eyes, in your style.
