Ollama's new app(ollama.com)

    Vendor lock-in: AFAIK it now uses a proprietary llama.cpp fork and builts its own registry on ollama.com in a kind of docker way (I heard docker ppl are actually behind ollama) and it's a bit difficult to reuse model binaries with other inference engines due to their use of hashed filenames on disk etc.

    Closed-source tweaks: Many llama.cpp improvements haven’t been upstreamed or credited, raising GPL concerns. They since switched to their own inference backend.

    Mixed performance: Same models often run slower or give worse outputs than plain llama.cpp. Tradeoff for convenience - I know.

    Opaque model naming: Rebrands or filters community models without transparency, biggest fail was calling the smaller Deepseek-R1 distills just "Deepseek-R1" adding to a massive confusion on social media and from "AI Content Creators", that you can run "THE" DeepSeek-R1 on any potato.

    Difficult to change Context Window default: Using Ollama as a backend, it is difficult to change default context window size on the fly, leading to hallucinations and endless circles on output, especially for Agents / Thinking models.

    llama.cpp: Battle-tested C++ engine with minimal deps and faster with many optimizations

    ik_llama.cpp: High-perf fork, even faster than default llama.cpp

    llama-swap: YAML-driven model swapping for your endpoint.

    LM Studio: GUI for any GGUF model—no proprietary formats with all llama.cpp optimizations available in a GUI

    Open WebUI: Front-end that plugs into llama.cpp, ollama, MPT, etc.