TUD:AI LLM API
An OpenAI-compatible LLM inference API operated by ZIH and ScaDS.AI Dresden/Leipzig for TU Dresden staff, students and ScaDS.AI affiliates, served from a LiteLLM proxy on the university's own network. Per TU Dresden's documentation it fronts chat, audio generation and transcription, image generation, embedding and reranking models, addressed through model aliases for code, high-availability, image-generation, reasoning and vision. The published context-length table names nine chat models: deepseek-ai/DeepSeek-V4-Flash and zai-org/GLM-5.3-Flash at 1,048,576 tokens, MiniMaxAI/MiniMax-M3, MiniMaxAI/MiniMax-M3-MXFP8 and zai-org/GLM-5.2-FP8 at 524,288, google/gemma-4-26B-A4B-it and Qwen/Qwen3.8-27B at 262,144, openai/gpt-oss-120b at 131,072 and meta-llama/Llama-3.3-70B-Instruct at 65,536. Keys minted through the Self-Service Portal carry a restrictive requests-per-minute default, raised on request. Verified live from outside the institution: GET /v1/models returns a LiteLLM auth_error envelope under HTTP 500 with no credential, and a materially different envelope naming a key hash when an invalid bearer token is presented, so the token is checked against a store. There is also a browser front end at chat.llm.scads.ai behind TUD account login. No OpenAPI is saved for this surface: the only contract served is LiteLLM's own, and crediting TU Dresden with it would repeat exactly the error this re-profile removed. This is the most consequential find of the pass — a German technical university running its own agent-facing inference API on its own hardware.