totally respect your choice, and it's a great project too. Of course as a maintainer of Ollama, my preference is to win you over with Ollama. If it doesn't meet your needs, it's okay. We are more energized than ever to keep improving Ollama. Hopefully one day we will win you back.
Ollama does not use llama.cpp anymore; we do still keep it and occasionally update it to remain compatible for older models for when we used it. The team is great, we just have features we want to build, and want to implement the models directly in Ollama. (We do use GGML and ask partners to help it. This is a project that also powers llama.cpp and is maintained by that same team)
Sorry, but this is kind of hiding the ball. You don't use llama.cpp, you just ... use their core library that implements all the difficult bits, and carry a patchset on top of it?
Why do you have to start with the first statement at all? "we use the core library from llama.cpp/ggml and implement what we think is a better interface and UX. we hope you like it and find it useful."
thanks, I'll take that feedback, but I do want to clarify that it's not from llama.cpp/ggml. It's from ggml-org/ggml. I supposed it's all interchangeable though, so thank you for it.
i.e. as of time of writing +/- 1445 lines between the two, on about 175k total lines. a lot of which is the recent MXFP4 stuff.
Ollama is great software. It's integral to the broader diffusion of LLMs. You guys should be incredibly proud of it and the impact its had. I understand the current environment rewards bold claims, but the sense I get from some of your communications is "what's the boldest, strongest claim we can make that's still mostly technically true". As a potential user, taking those claims as true until closer evaluation reveals the discrepancy feels pretty bad, and keeps me firmly in the 'potential' camp.
Have the confidence in your software and the respect for your users to advertise your system as it is.
I'm torn on this, I was a fan of the project from the very beginning and never sent any of my stuff upstream, so I'm less than a contributor but more than don't care, and it's still non-obvious how the split happened.
But the takeaway is pretty clearly that `llama.cpp`, `GGML`/`GGUF`, and generally `ggerganov`'s single-handedly Carmacking it when everyone thought it was impossible is all the value. I think a lot of people made Docker containers with `ggml`/`gguf` in them and one was like "we can make this a business if we realllllly push it".
Ollama as a hobby project or even a serious OSS project? With a cordial upstream relationship and massive attribution labels everywhere? Sure. Maybe even as a commercial thing that has a massive "Wouldn't Be Possible Without" page for it's OSS core upstream.
But like: startup company for making money that's (to all appearances) completely out of reach for the principles to ever do without totally `cp -r && git commit` repeatedly? It's complicated, a lot of stuff starts as a fork and goes off in a very different direction, and I got kinda nauseous and stopped paying attention at some point, but near as I can tell they're still just copying all the stuff they can't figure out how to do themselves on an ongoing basis without resolving the upstream drama?
It's like, in bounds barely I guess. I can't point to it being "this is strictly against the rules or norms", but it's bending everything to the absolute limit. It's not a zone I'd want to spend a lot of time in.
To be clear I was comparing ggml-org/ggml to ggml-org/llama.cpp/ggml to respond to the earlier thing. Ollama carries an additional patchset on top of ggml-org/ggml.
> [ggml] is all the value
That’s what gets me about Ollama - they have real value too! Docker is just the kernel’s cgroups/chroots/iptables/… but it deserves a lot of credit for articulating and operating those on behalf of the user. Ollama deserves the same. But they’re consistently kinda weird about owning just that?
So I’m using turbo and just want to provide some feedback. I can’t figure out how to connect raycast and project goose to ollama turbo. The software that calls it essentially looks for the models via ollama but cannot find the turbo ones and the documentation is not clear yet. Just my two cents, the inference is very quick and I’m happy with the speed but not quite usable yet.
I won't use `ollama` on principle. I use `llama-cli` and `llama-server` if I'm not linking `ggml`/`gguf` directly. It's like, two extra commands to use the one by the genius that wrote it and not the one that the guys just jacked it.
The models are on HuggingFace and downloading them is `uvx huggingface-cli`, the `GGUF` quants were `TheBloke` (with a grant from pmarca IIRC) for ages and now everyone does them (`unsloth` does a bunch of them).
Maybe I've got it twisted, but it seems to be that the people who actually do `ggml` aren't happy about it, and I've got their back on this.
I'm the first to admit I'm not a heavy C++ user, so I'm not a great judge of the quality looking at the code itself ... but ggml-org has 400 contributors on ggml, 1200 on llama.cpp and has kept pace with ~all major innovations in transformers over the last year and change. Clearly some people can and do make meaningful contributions.
Interesting, admittedly, I am slowly getting to the point, where ollama's defaults get a little restrictive. If the setup is not too onerous, I would not mind trying. Where did you start?
Download llama-server from llama.cpp Github and install it some PATH directory. AFAIK they don't have an automated installer, so that can be intimidating to some people
Assuming you have llama-server installed, you can download + run a hugging face model with something like
First, I must say I appreciate you taking the time to be engaged on this thread and responding to so many of us.
What I'm referring to is a broader pattern that I (and several) others have been seeing. Of the top of my head: not crediting llama.cpp previously, still not crediting llama.cpp now and saying you are using your own inference engine when you are still using ggml and the core of what Georgi made, most importantly why even create your own version - is it not better for the community to just contribute to llama.cpp?, making your own propreitary model storage platform disallowing using weights with other local engines requiring people to duplicate downloads and more.
I dont know how to regard these other than being largely motivated out of self interest.
I think what Jeff and you have built have been enormously helpful to us - Ollama is how I got started running models locally and have enjoyed using it for years now. For that, I think you guys should be paid millions. But what I fear is going to happen is you guys will go the way of the current dogma of capturing users (at least in mindshare) and then continually squeezing more. I would love to be wrong, but I am not going to stick around to find out as its risk I cannot take.
In an ideal world yes - as we should - especially for us Californian/Bay Area people, that's literally our spirit animal. But I understand that is idle dreaming. What I believe certainly is within reach is a state that is much better than what we are in.
It needn't be idle dreaming? What fundamental law or societal agreement prevents solarpunk versus the current status quo of corporate anti-human cyberpunk?
Yes, better to get free sh*t unsustainably. By the way, you're free to create an open source alternative and pour your time into that so we can all benefit. But when you don't — remember I called it!
What? The obvious move is to never have switched to Ollama and just use Llama.cpp directly, which I've been doing for years. Llama.cpp was created first, is the foundation for this product, and is actually open source.
I do also need an API server though. The one built into OpenWebUI is no good because it always reloads the model if you use it first from the web console and then run an API call using the same model (like literally the same model from the workspace). Very weird but I avoid it for that reason.
llama.cpp is what you want. It offers both a web UI and an API on the same port. I use llama.cpp's webui with gpt-oss-20b, and I also leverage it as an OpenAI-compatible server with gptel for Emacs. Very good product.
Most apps that integrate with ollama that I've seen just have an OpenAI compatible API parameter which defaults to port 11434 which ollama uses, but can be changed easily. Is there a way to integrate ollama more deeply?
It's very unfortunate that the local inference community has aggregated around Ollama when it's clear that's not their long term priority or strategy.
Its imperative we move away ASAP