My favorite LM Studio alternative isn't Ollama, it's something faster and more lightweight

Local LLMs have become an actual daily part of a lot of workflows, and the runner space has grown fast alongside it. LM Studio is still the easiest first runner for anyone getting into local AI in my opinion, with Ollama being a close second. But they're not the only options worth looking at, and they certainly have their own shortcomings that makes it worth digging around.

KoboldCpp is the one that kept coming up in self-hosting circles for a while before I actually tried it. It's built on the same llama.cpp core that LM Studio and Ollama are wrappers around, but the way it packages everything is different, and the performance difference on the same hardware is pretty notable too.

The runner that packs everything into one file

Small download but massive feature set

KoboldCpp is an open-source single-file executable for running local LLMs, and it comes with its own embedded web UI called Kobold Lite. That's what you actually chat with once a model is loaded. The first beta released in March 2023, developed by LostRuins on GitHub, which actually puts it a bit earlier than most of the local runners people talk about now. The dev's whole design philosophy is captured in the repo tagline: "One File. Zero Install". Meaning there's no Docker Python setup, CUDA toolkit install, or npm, which is a big part of why I think many bounce off local LLM tooling in the first place.

The feature set is honestly wider than the "text generation runner" label suggests. GGUF text models are the main thing, but Kobold also does image recognition via mmproj files, image generation through a bundled Stable Diffusion.cpp, speech-to-text via Whisper, text-to-speech through several TTS engines, and even music generation. And it exposes OpenAI-compatible, Ollama-compatible, and A1111 endpoints at the same time, so it slots into whatever downstream tools you're already using.

Where KoboldCPP leaves the others behind

Speed is the biggest difference

Speed is where I noticed the difference first, and it's bigger than I thought. On the same Gemma 4 E4B model, same 12k context, and same GPU obviously, LM Studio pulled around 11 tok/sec on longer generations. Kobold on the exact same setup pushed closer to 80. That's not margin of error, it's the same output coming through in a fraction of the time. I've re-run this a few times just to make sure I wasn't reading it wrong.

Then there's audio input, which Kobold picks up automatically when you load a compatible mmproj. Gemma 4's vision projector actually includes audio hparams too, which I only realized when the terminal started listing them during load. LM Studio doesn't do audio input yet, so if your model has that mode you're missing out entirely.

Memory is another one worth mentioning. Kobold Lite has Memory, Author's Note, and AutoGenerate Memory, and they're all editable in the same panel. LM Studio has nothing native for this. There are community plugins available for this, but it's a plugin, and you have to know it exists to go find it, then set it up. It's always nicer when a tool just ships with everything you need.

ContextShift is the last one, and it might be the win of them all. When your context window fills up in LM Studio, for example, it reprocesses the whole thing from scratch, which is what causes that pause between long responses. Kobold shifts the existing context instead of rebuilding it, so there's no need to reprocess it. For long chats, this is a very appreciated time-saver.

The launcher shows you what's actually running

Nothing is hidden behind a settings menu

One thing I love about Kobald is when you open it the Quick Launcher puts everything in front of you at once. Backend selection with CUDA, Vulkan, or CPU, your GPU auto-detected by name, GPU layer count, MMQ, FlashAttention, ContextShift, MMAP, a Context Size slider, the GGUF path field, and much more (half of these I don't even need or fully understand, but they're there if you need them).

LM Studio splits the same controls across the model load flyout and a separate settings panel, and Ollama doesn't surface most of them at all outside of environment variables and Modelfile edits. FlashAttention is also in a different place than GPU offload in LM Studio and behind the shell export in Ollama, so you never really see the full picture of how your model is about to run.

I think having all these controls is an advantage regardless of who's using the tool. Power users get to tune without hunting through menus, and beginners can ignore most of it and just hit Launch because the defaults are pretty sensible. This setup works because Kobold is just one file wrapping around llama.cpp with nothing between you and it.

It does what the other runners do and more (and faster)

KoboldCPP isn't going to replace LM Studio or Ollama for everyone, and I'll probably keep LM Studio in rotation just because I'm so used to it at this point. It's just a different answer to the same problem, aimed at people who want the runner to be comprehensive without being heavy, and actually expose what llama.cpp is doing.