You can compile it from source, all you need to do is clone the repository and do a `cmake -B build -DGGML_VULKAN=1` (add other backends if you want) followed by a `cmake --build build --config Release` and then you get all the llama tools in the `build/bin` (including `llama-server` which provides a web-based interface). There is a `docs/build.md` that has more detailed info (especially if you need another backend, though at least on my RX 7900 XTX i see no difference in terms of performance between Vulkan and ROCm and the former is much more stable and compatible -- i tried ROCm for a bit thinking it'd be much faster but only ended up being much more annoying as some models would OOM on it while they worked on Vulkan -- if you or NVIDIA hardware all this may sound quaint though :-P).

Why are you looking to move off Ollama? Just curious because I'm using Ollama and the cloud models (Kimi 2.5 and Minimax 2.7) which I'm having lots of good success with.