Honestly you just need koboldcpp. It is one click and supports everything and is Foss
I started off with Ollama, but liked llama.cpp better because you can download one GGUF file the day it’s released and add it to a source controlled config.ini with parameters to make it work best on your hardware.
I didn’t get serious about local LLMs until I realized I could run Qwen 3.6 35B-A3B on my 3070 with q4 for KV cache and get over 100k context. I found Ollama’s Modelfile approach and model handling in general to be clunky and discouraging important tweaks, but llama.cpp keeps everything simple.
The article feels like an LLM helped write it, but it’s comprehensive nonetheless. I recently switched to bare llama.cpp since I have limited hardware capabilities and this squeezes a tiny bit more out of it. However, my usage is limited to learning / testing
The article feels like an LLM helped write it
Nah. It’s got a ‘millennial’ amount of run-on sentences and comma-splice issues; stuff an LLM wouldn’t do.
So far I’ve been using a wrapper for wrappers called ramalama which makes opinionated choices based on your hardware. My main machine is a 64 core Arm machine with an 8gb AMD card using Vulkan. I’m sure I could do better but it’s a lot of tweaking I don’t have time to learn about right now.
lm studio here
I’m using Koboldcpp for gguf models I can get from anywhere. I feel like I’m missing some other obvious choices for me.
I switched to LM Studio as my generic “run whatever GGUF I’ve come across today” platform and I quite like it. It doesn’t have the weird setup and configuration thing Ollama does, it’s got a nice built-in chat system like Kobold does, and it does dynamic model loading on demand so my various random scripts can call into it for whatever they need without me having to do any manual work.
Running stuff with straight llama would probably be more computationally efficient, but I value easy interfaces too.
Started with llama file and run vllm now as a container.
