I prefer slow consolidated updates than a rat race on the newest personally, but i do nothing enough critical to depend on the last nightly llama.cpp. So it's biased i guess. By the way you can select the branch you want in the settings, if you like to run on glowing embers \^\^ Force-update too. And others stuffs like this.

For the OSS part, it's doomed even if the team is quite open to talk, wich is quite rare in the game now.

Now if you find a llama.cpp work that run MCPs, LM Link clone, Plugins, Rag, that is barebone ... i'll jump. The only thing that i dislike in LMS is the hidden sysprompts for users using it as main chat interface and not as an inferences services. The bloat is not big but this layer is disturbing for little models and a total waste of compute power for local models; the prompts are not even systems or a good code but useless prose "you are a smart assistant, blablabla ...". I'm astonished that a bunch of respectfull companies are still RPing their products today, make no sense for me with the backward we have now in 2026 on what is really happening in the latent space of LLMs \^\^ Even very contradictory with the maths behind this lol