The migration resistance has a concrete number: Ollama is treated as finished at about five concurrent users, which is what pushes teams onto vLLM or SGLang

The most operationally useful sentiment datum in the corpus is an explicit concurrency ceiling given by someone running a real organization on it. A single-firm deployment (Threadripper Pro 7965WX, 256GB DDR5-6000, one L40S, roughly EUR 14k, Docker on native Ubuntu, OpenWebUI for colleagues) states: 'For engine go with your flavour, Ollama is good enough for 5 active concurrent users but I wouldn't go futher than that. Simply vllm or sglang or aphrodite or llamacpp is all you need.' The same commenter independently validates the engine-selection split from the deployment side: 'For engine go with your flavour' - i.e. no strong preference at that scale. Note this is a preference statement carrying a number, not a benchmark of that number; the number is a practitioner rule of thumb.

Ollama optimizes for a single developer's machine - one model, one user, automatic configuration - and has no continuous-batching server behind it. Past a handful of concurrent clients the choice is not 'which is faster' but 'which has a scheduler'. This is the single most common named trigger for moving off a simple local stack, and it is why the migration in this corpus is usually Ollama to vLLM rather than engine to engine.

Below roughly five concurrent users and a single model, a simple single-process stack is the corpus's recommendation and no datacenter engine is warranted. Above it, move to an engine with continuous batching. Treat the number five as a practitioner rule of thumb to be measured on your own traffic, not a benchmark result.

Same deployment, one further preference worth recording because it is about the frontend rather than the engine: OpenWebUI was chosen for colleagues 'due to how simple it is for the end user, not to complicated and such like other front ends yet having a lot of features.' And one responder to 'which GPU should I use to caption ~50k images/day' territory is worth noting as a general shape: batch-oriented, latency-insensitive work attracts different engine advice than interactive work. Recorded as a caveat on the ceiling's scope, not as a benchmark.