University Medicine Magdeburg
Magdeburg university hospital hosts Llama 3.2-Vision on its own server for 322 staff
University Medicine Magdeburg in Germany set up a self-hosted AI chat platform on Meta's open-weight Llama 3.2-Vision 90B, running on one server it owns. In a staff survey two months after launch, 70.3% of respondents rated its reliability and speed as high.
Key figures
Copied exactly from the sources. Copy any figure with its citation.
44%
(40 people)
Respondents who reported saving at least 30 minutes of work per week, per the paper's abstract (self-reported)1
What it took
Cost, team size and timeline were not disclosed in the sources.
The problem
Hospital staff were already using public AI chatbots for writing and information searches. Hospitals handle sensitive personal data, and sending it to an outside service can conflict with data protection rules. The Data Integration Center at University Medicine Magdeburg wanted a tool staff could use without data leaving the building.
What they built
The center built an AI chat platform on Llama 3.2-Vision 90B, an open-weight model from Meta. Open-weight means the model's files are published, so an organisation can run the model on its own hardware. The model can read text and images. The platform launched on 3 March 2025 for administration, health care and research staff.
It runs on one server the hospital owns, with an NVIDIA H100 graphics chip (80 GB of memory). The paper reports 25.3 output tokens per second. A token is a small piece of a word.
How they did it
The team announced the platform through the hospital's circular mailing list, and staff registered with their organisational unit. Two months after launch it sent a ten-question online survey to registered users, open for 26 days with two reminders. It asked about frequency of use, tasks, satisfaction and weekly time saved. The study was exempt from ethics approval because it involved no intervention or sensitive information.
Results
Of 322 registered users, 98 took part in the survey and 91 gave complete answers. The paper's abstract reports that 44% (40 people) said they saved at least 30 minutes of work per week. Reliability and speed were rated high by 70.3%, and answer quality by 60.4%. Most respondents (85.7%) needed under 15 minutes to get started.
Use differed by job. Health care and research staff used the platform more often, and for creation and analysis tasks, than administrators.
What went wrong
The most common criticism, from 37.4% of respondents, concerned the quality of answers. About one third of those comments (12 of 34) mentioned false answers, such as invented references. Respondents also asked for file upload, export options, documentation and training, and integration with the hospital information system.
What to copy
Run a short survey soon after launch, and use it to decide training, features and which staff groups to target next.
All reported figures
| Measure | Figure | Citation |
|---|---|---|
| Registered users of the platform | 322 (21.7% of the 1,484 employees in the target group)1 | |
| Survey respondents with complete answers | 91 of 98 who took part1 | |
| Respondents who reported saving at least 30 minutes of work per week, per the paper's abstract (self-reported) | 44% (40 people)1 | |
| Respondents who rated reliability and speed as extremely or moderately high | 70.3% (64 of 91)1 | |
| Respondents who rated answer quality as extremely or moderately high | 60.4% (55 of 91)1 | |
| Respondents who needed under 15 minutes to get familiar with the platform | 85.7% (78 of 91)1 | |
| Free-text criticism about the quality of answers | 37.4% (34 of 91)1 |
Lessons for your project
A single mid-sized server can serve a hospital pilot. This one used one NVIDIA H100 chip and produced 25.3 output tokens (word pieces) per second.
Expect uneven adoption by role. Health care and research staff used the platform more than administrators, so target training at the groups that need it.
Measure perceived value early. A short survey two months after launch showed where users saw time savings and what they wanted next.
Plan for hallucinations. The most common complaint was wrong answers, including invented references, so tell users to check sources.
Caveats
- The authors are the hospital's own data team, and the time savings are self-reported by volunteer survey respondents (91 of 322 registered users).
- Only 30.4% of registered users answered the survey, so results reflect self-selected respondents. The authors did not run statistical tests because the groups were small.
- The paper does not state a project budget or team size.
- The 44% time-saving figure appears only in the abstract. The results section does not restate it for the whole sample, and the same 40 of 91 numbers appear there for administration staff, so treat it with care.
- The paper gives the share of weekly users as 62.6% (57) in the abstract and 61.5% (56) in the text, so this report omits it.
Sources
- Implementation and User Evaluation of an On-Premise Large Language Model in a German University Hospital Setting: Cross-Sectional Survey
JMIR AI · 2026-04-15 · Research paper · ai.jmir.org
Summarized in our own words from the sources above. Spotted an error? See corrections.