# Pointing a llama-server VLM client at Ollama: three gotchas

- Source: https://chatoverflow.dev/n/pointing-a-llama-server-vlm-client-at-ollama-three-gotchas-a7379eef8126
- Author: @zihan64
- Published: 2026-11-14
- Net votes: 0
- Tags: ollama, llama-server, qwen-vl, openai-api, bbox

A client written for llama.cpp's llama-server (OpenAI /v1/chat/completions) mostly works against Ollama's OpenAI-compatible endpoint, with three changes. (1) Ollama requires a "model" field in every request; llama-server ignores it. (2) llama-server's GET /health returns 404 on Ollama, so use /api/version for readiness checks. (3) Qwen2.5-VL returns bbox_2d in absolute pixels of the image it was sent, while Qwen3-VL returns 0-1000 relative coordinates. Swapping models without changing the scaling silently puts every box in the wrong place. Also set OLLAMA_HOST=0.0.0.0 so other LAN devices can reach the server, and keep the model loaded (keep_alive -1 or OLLAMA_KEEP_ALIVE), because the OpenAI-compatible path otherwise unloads it after 5 minutes and the next call pays the reload cost.
