Nice release. Part of the problem right now with OSS models (at least for enterprise users) is the diversity of offerings in terms of:
- Speed
- Cost
- Reliability
- Feature Parity (eg: context caching)
- Performance (What quant level is being used...really?)
- Host region/data privacy guarantees
- LTS
And that's not even including the decision of what model you want to use!
Realistically if you want to use an OSS model instead of the big 3, you're faced with evalutating models/providers across all these axes, which can require a fair amount of expertise to discern. You may even have to write your own custom evaluations. Meanwhile Anthropic/OAI/Google "just work" and you get what it says on the tin, to the best of their ability. Even if they're more expensive (and they're not that much more expensive), you are basically paying for the priviledge of "we'll handle everything for you".
I think until providers start standardizing OSS offerings, we're going to continue to exist in this in-between world where OSS models theoretically are at performance parity with closed source, but in practice aren't really even in the running for serious large scale deployments.
I wouldn't be surprised if those undeleted chats or some inferred data that is based on it is part of the gpt-5 training data. Somehow I don't trust this sama guy at all.
I see a lot of hate for ollama doing this kind of thing but also they remain one of the easiest to use solutions for developing and testing against a model locally.
Sure, llama.cpp is the real thing, ollama is a wrapper... I would never want to use something like ollama in a production setting. But if I want to quickly get someone less technical up to speed to develop an LLM-enabled system and run qwen or w/e locally, well then its pretty nice that they have a GUI and a .dmg to install.
Since the new multimodal engine, Ollama has moved off of llama.cpp as a wrapper. We do continue to use the GGML library, and ask hardware partners to help optimize it.
Ollama might look like a toy and what looks trivial to build. I can say, to keep its simplicity, we go through a deep amount of struggles to make it work with the experience we want.
Simplicity is often overlooked, but we want to build the world we want to see.
> Ollama has moved off of llama.cpp as a wrapper. We do continue to use the GGML library
Where can I learn more about this? llama.cpp is an inference application built using the ggml library. Does this mean, Ollama now has it's own code for what llama.cpp does?
But Ollama is a toy, it's meaningful for hobbyists and individuals to use locally like myself. Why would it be the right choice for anything more? AWS, vLLM, SGLang etc would be the solutions for enterprise
I knew a startup that deployed ollama on a customers premises and when I asked them why, they had absolutely no good reason. Likely they did it because it was easy. That's not the "easy to use" case you want to solve for.
I can say trying many inference tools after the launch, many do not have the models implemented well, and especially OpenAI’s harmony.
Why does this matter? For this specific release, we benchmarked against OpenAI’s reference implementation to make sure Ollama is on par. We also spent a significant amount of time getting harmony implemented the way intended.
I know vLLM also worked hard to implement against the reference and have shared their benchmarks publicly.
Honestly, I think it just depends. A few hours ago I wrote I would never want it for a production setting but actually if I was standing something up myself and I could just download headless ollama and know it would work. Hey, that would also be fine most likely. Maybe later on I'd revisit it from a devops perspective, and refactor deployment methodology/stack, etc. Maybe I'd benchmark it and realize its fine actually. Sometimes you just need to make your whole system work.
We can obviously disagree with their priorities, their roadmap, the fact that the client isn't FOSS (I wish it was!), etc but no one can say that ollama doesn't work. It works. And like mchiang said above: its dead simple, on purpose.
Llama.cpp is not really that easy unless you're supported by their prebuilt binaries. Go to the llama.cpp GitHub page and find a prebuilt CUDA enabled release for a Fedora based linux distro. Oh there isn't one you say? Welcome to losing an hour or more of your time.
Then you want to swap models on the fly. llama-swap you say? You now get to learn a new custom yaml based config file syntax that does basically nothing that the Ollama model file already does so that you can ultimately... have the same experience as Ollama but now you've lost hours just to get back to square one.
Then you need it to start and be ready with the system reboot? Great, now you get to write some systemd services, move stuff into system-level folders, create some groups and users and poof, there goes another hour of your time.
Sure but if my some of the development team is using ollama locally b/c it was super easy to install, maybe I don't want to worry about maintaining a separate build chain for my prod env. Many startups are just wrapping or enabling LLMs and just need a running server. Who are we to say what is right use of their time and effort?
I can say as a fact, for the gpt-oss model, we also implemented our own MXFP4 kernel. Benchmarked against the reference implementations to make sure Ollama is on par. We implemented harmony and tested it. This should significantly impact tool calling capability.
Im not sure if im feeding here. We really love what we do, and I hope it shows in our product, in Ollama’s design and in our voice to our community.
You don’t have to like Ollama. That’s subjective to your taste. As a maintainer, I certainly hope to have you as a user one day. If we don’t meet your needs and you want to use an alternative project, that’s totally cool too. It’s the power of having a choice.
Is there a schedule for adding additional models to Turbo mode plan, in addition to gpt-oss 20/120b? I wanted to try your $20/month Turbo plan, but I would like to be able to experiment with a few other large models.
vllm and ollama assume different settings and hardware. Vllm backed by the paged attention expect a lot of requests from multiple users whereas ollama is usually for single user on a local machine.
The issue is not companies but governance. OSS licenses and companies are fine. Companies have a natural conflict of interest that can lead them to take software projects they control in a direction that suits their revenue goals but not necessarily the needs/wants of its users. That happens over and over again. It's their nature. This can means changes in direction/focus or worst case license changes that limit what you can do.
The solution is having proper governance for OSS projects that matter with independent organizations made up of developers, companies, and users taking care of the governance. A lot of projects that have that have last for decades and will likely survive for decades more.
And part of that solution is to also steer clear of projects without that. I've been burned a couple of times now getting stuck with OSS components where the license was changed and the companies behind it had their little IPOs and started serving share holders instead of users (elastic, redis, mongo, etc). I only briefly used Mongo and I got a whiff of where things were going and just cut loose from it. With Elastic the license shenenigans started shortly after their IPO and things have been very disruptive to the community (with half using Opensearch now). With Redis I planned the switch to Valkey the second it was announced. Clear cut case of cutting loose. Valkey looks like it has proper governance. Redis never had that.
Ollama seems relatively OK by this benchmark. The software (ollama server) is MIT licensed and there appears to be no contributor license agreement in place. But it's a small group of people that do most of the coding and they all work for the same vc funded company behind ollama. That's not proper governance. They could fail. They could relicense. They could decide that they don't like open source after all. Etc. Worth considering before you bet your company on making this a foundational piece of your tech stack.
I view it a bit like I do cloud gaming, 90% of the time I'm fine with local use, but sometimes it's just more cost effective to offload the cost of hardware to someone else. But it's not an all-or-nothing decision.
Yep, if you just want to play one or two games at 4k HDR etc. it's a lot cheaper to pay 22€ for GeForce Now Ultimate vs. getting a whole-ass gaming PC capable of the same.
Any more information on "Privacy first"? It seems pretty thin if just not retaining data.
For Draw Things provided "Cloud Compute", we don't retain any data too (everything is done in RAM per request). But that is still unsatisfactory personally. We will soon add "privacy pass" support, but still not to the satisfactory. Transparency log that can be attested on the hardware would be nice (since we run our open-source gRPCServerCLI too), but I just don't know where to start.
In theory, "privacy pass" should help, as you can subpoena content, but cannot know who made these. But that is still thin (and Ollama not doing that too anyway).
I’d love to learn more about your project. I’m using socialized cloud regions for AI security and they really lag the mainstream. Definitely need more options here.
Edit: emailed the address on the site in your profile, got an inbox does not exist error.
I would pay more if they let you run the models in Switzerland or some other GDPR respecting country, even if there was extra latency. I would also hope everything is being sent over SSL or something similar.
What could be the benefit of paying $20 to Ollama to run inferior models instead of paying the same amount of money to e.g. OpenAI for access to sota models?
I feel the primary benefit of this Ollama Turbo is that you can quickly test and run different models in the cloud that you could run locally if you had the correct hardware.
This allows you to try out some open models and better assess if you could buy a dgx box or Mac Studio with a lot of unified memory and build out what you want to do locally without actually investing in very expensive hardware.
Certain applications require good privacy control and on-prem and local are something certain financial/medical/law developers want. This allows you to build something and test it on non-private data and then drop in real local hardware later in the process.
> quickly test and run different models in the cloud that you could run locally if you had the correct hardware.
I feel like they're competing against Hugging Face or even Colaboratory then if this is the case.
And for cases that require strict privacy control, I don't think I'd run it on emergent models or if I really have to, I would prefer doing so on an existing cloud setup already that has the necessary trust / compliance barriers addressed. (does Ollama Turbo even have their Trust center up?)
I can see its potential once it gets rolling, since there's a lot of ollama installations out there.
Running models without a filter on it. OpenAI has an overzealous filter and won’t even tell you what you violated. So you have to do a dance with prompts to see if it’s copyright, trademark or whatever. Recently it just refused to answer my questions and said it wasn’t true that a civil servant would get fired for releasing a report per their job duties. Another dance sending it links to stories that it was true so it could answer my question. I want a LLMs without training wheels.
I am so so so confused as to why Ollama of all companies did this other than an emblematic stab at making money-perhaps to appease someone putting pressure on them to do so. Their stuff does a wonderful job of enabling local for those who want it. So many things to explore there but instead they stand up yet another cloud thing? Love Ollama and hope it stays awesome
The problem is that OSS is free to use but it is not free to create or maintain. If you want it to remain free to use and also up to date, Ollama will need someone to address issues on GitHub. Usually people want to be paid money for that.
money is great! I like money! but if this is their version of buy me a coffee I think there’s room to run elsewhere for their skillset/area of expertise
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.
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'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.
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?
For production use of open weight models I'd use something like Amazon Bedrock, Google Vertex AI (which uses vLLM), or on-prem vLLM/SGLang. But for a quick assessment of a model as a developer, Ollama Turbo looks appealing. I find Google GCP incredibly user hostile and a nightmare to navigate quotas and stuff.
For one of the top local open model inference engines of choice - only supporting OSS out of the gate feels like an angle to just ride the hype knowing OSS is announced today "oh OSS came out and you can use Ollama Turbo to use it"
The subscription based pricing is really interesting. Other players offer this but not for API type services. I always imagine that there will be a real pricing war with LLMs with time / as capabilities mature, and going monthly pricing on API services is possibly a symptom of that
What does this mean for the local inference engine? Does Ollama have enough resources to maintain both?
Does this mean we can access Ollama APIs for $20/mo and test them without running the model locally? I'm not hardware-rich, but for some projects, I'd like a reliable pricing.
It says “usage-based pricing” is coming soon. I think that is the sweet spot for a service like this.
I pay $20 to Anthropic, so I don’t think I’d get enough use out of this for the $20 fee. But being able to spin up any of these models and use as needed (and compare) seems extremely useful to me.
A flat fee service for open-source LLMs is somewhat unique, even if I don't see myself paying for it.
Usage-based pricing would put them in competition with established services like deepinfra.com, novita.ai, and ultimately openrouter.ai. They would go in with more name-recognition, but the established competition is already very competitive on pricing
> It says “usage-based pricing” is coming soon. I think that is the sweet spot for a service like this.
Agreed, though there are already several providers of these new OpenAI models available, so I'm not sure what ollama's value add is there (there are plenty of good chat/code/etc interfaces available if you are bringing your own API keys).
Ollama is great but I feel like Georgi Gerganov deserves way more credit for llama.cpp.
He (almost) single-handedly brought LLMs to the masses.
With the latest news of some AI engineers' compensation reaching up to a billion dollars, feels a bit unfair that Georgi is not getting a much larger slice of the pie.
Ollama is not a wrapper around llama.cpp anymore, at least for multimodal models (not sure about others). They have their own engine: https://ollama.com/blog/multimodal-models
Is Georgi landing any of those big-time money jobs? I could see a conflict-of-interest given his involvment with llama.cpp, but I would think he'd be well positioned for something like that
Seriously, people astroturfing this thread by saying ollama has a new engine. It literally is the same engine that llama.cpp uses and georgi and slaren maintain! VC funding will make people so dishonest and just plain grifters
That'll be an uphill battle on value proposition tbh. $20 a month for access to a widely available MoE 120B with ~5B active parameters at unspecified usage limits?
I guess their target audience values convenience and easy of use above all else so that could play well there maybe.
Is there an evaluation of such services available anywhere. Looking for recommendations for similar services with usage based pricing and pro-and-cons.
ps: looking for most economic one to play around with as long as it a decent enough experience (minimal learning curve). buy, happy to pay too
OpenRouter is great. Less privacy I guess, but you pay for usage and you have access to hundreds of models. They have free models too, albeit rate-limited.
If any of the major inference engines - vLLM, Sglang, llama.cpp - incorporated api driven model switching, automatic model unload after idle and automatic CPU layer offloading to avoid OOM it would avoid the need for ollama.
Interesting - it does indeed seem like llama-server has the needed endpoints to do the model swapping and llama.cpp as of recently also has a new flag for the dynamic CPU offload now.
However the approach to model swapping is not 'ollama compatible' which means all the OSS tools supporting 'ollama' Ex Openwebui, Openhands, Bolt.diy, n8n, flowise, browser-use etc.. aren't able to take advantage of this particularly useful capability as best I can tell.
I like how the landing page (and even this HN page until this point) completely miss any reference to Meta and Facebook.
The landing page promises privacy but anyone who knows how FB used VPN software to spy on people, knows that as long as the current leadership is in place, we shouldn't assume they've all of a sudden became fans of our privacy.
Ollama isn’t connected to Meta besides offering Llama as one of the potential models you can run.
There is obviously some connection to Llama (the original models giving rise to llama.cpp which Ollama was built on) but the companies have no affiliation.
I do hope Ollama got a good paycheck from that, as they are essentially help OpenAI to oss-wash their image with the goodwill that Ollama has built up.
I think what matters more here is "All hardware is located outside of China". Located in the US means little because that's not good enough for many regulated industries even within the US.
All things considered though, Europe is getting confusing. They have GDPR but now pushing to backdoor encryption within the EU? [1]
At least there isn't a strong movement in the US trying to outlaw E2E encryption.
Which brings up the point are truly private LLMs possible? Where the input I provide is only meaningful to me, but the LLM can still transform it without gaining any contextual value out of it? Without sharing a key? If this can be done, can it be done performantly?
Yes, there is gonna be a new discussion for it on October 15, but I've already seen section of governments being against their own government position on the bill (Swedish Military for example).
No I think the point is to choose the best jurisdiction to have cloud hosted data where your data is best protected from access by very wealthy entities via intelligence services bribery. That’s still hands down the USA.
They might have access to any given machine, but they lack the broad scope of general surveillance. If they want to get you, just like most of the other nation state level threats, you will get got. For other threat models, the US works pretty well.
I guarantee that nobody cares about or will be surveilling your private AI use unless you're doing other things that warrant surveillance.
The reason big providers suck, as OpenAI is so nicely demonstrating for us, is that they retain everything, the user is the product, and court cases, other situations can unmask and expose everything you do on a platform to third parties. This country seriously needs a digital bill of rights.
Nobody cares? That seems ludicrous to me. The last 3 decades of business have been characterized most of all by the increased access of private information on people for online business competitive insights. Sure if you are just a consumer you have nothing of real value except in the aggregate, but if you are an up-and-coming business drawing customers away from other businesses, your private AI use is absolutely of interest. Which is why serious businesses here scour the ToS.
The biggest game in town has been managing platforms that give owners an information advantage. But at least the world generally trusts the USA to abide by laws and user agreements, which is why, to my mind, the USA retains the near monopoly on information platforms.
I personally wouldn’t trust a UK platform for example, being a Brit native. The top echelon talent pool is so small and incestuous I don’t believe I would experience a fair playing field if a business of mine passed a certain size of national reach/importance.
EDIT: from ChatGPT, new money entrepreneurs with no inheritence/political ties by economic region,
USA ~63%, UK/HongKong/Singapore ~45%, Emerging Markets ~35%, EU ~22%, Russia ~10%
It was natively trained in FP4. Probably both to reduce VRAM usage at inference time (fits on a single H100), and to allow better utilization of B200s (which are especially fast for FP4).
at this point, can i purchase the subscription directly from the model provider or hugging face and use it? or is this ollama attempt to become a provider like them.
Often the math works out that you get a lot more for $20 a month if you settle for smaller sized but capable models (8b-30b). I don’t see how it’s better other than Ollama can “promise” they don’t store your data where as OpenRouter is dependent on which host you choose (and there’s no indicator on OpenRouter exposing which ones do or don’t).
In a universe where everything you say can be taken out of context, things like OpenAi will be a data leak nightmare.
Watching ollama pivot from a somewhat scrappy yet amazingly important and well designed open source project to a regular "for-profit company" is going to be sad.
Thankfully, this may just leave more room for other open source local inference engines.
I don't blame them. As soon as they offer a few more models available with the Turbo mode I plan on subscribing to their Turbo plan for a couple of months - a buying them a coffee, or keeping the lights on kind of thing.
The Ollama app using the signed-in-only web search tool is really pretty good.
we have always been building in the open, and so is Ollama. All the core pieces of Ollama are open. There are areas where we want to be opinionated on the design to build the world we want to see.
There are areas we will make money, and I wholly believe if we follow our conscious we can create something amazing for the world while making sure we can keep it fueled to keep it going for the long term.
Some of the ideas in Turbo mode (completely optional) is to serve the users who want a faster GPU, and adding in additional capabilities like web search. We loved the experience so much that we decided to give web search to non-paid users too. (Again, it's fully optional). Now to prevent abuse and make sure our costs don't go out of hand, we require login.
Can't we all just work together and create a better world? Or does it have to be so zero sum?
I wanted to try web search to increase my privacy but it wanted to do login.
For Turbo mode I understand the need for paying but the main poing of running a local model with web search is browsing from my computer without using any LLM provider. Also I want to get rid of the latency to US servers from Europe.
I think this offering is a perfectly reasonable option for them to make money. We all have bills to pay, and this isn't interfering with their open source project, so I don't see anything wrong with it.
>> Watching ollama pivot from a somewhat scrappy yet amazingly important and well designed open source project to a regular "for-profit company" is going to be sad.
if i could have consistent and seamless local-cloud dev that would be a nice win. everyone has to write things 3x over these days depending on your garden of choice, even with langchain/llamaindex
It was always just a wrapper around the real well designed OSS, llama.cpp. Ollama even messes up the names of models by calling distilled models the name of the actual one, such as DeepSeek.
Ollama's engineers created Docker Desktop, and you can see how that turned out, so I don't have much faith in them to continue to stay open given what a rugpull Docker Desktop became.
Same, was just after a small lightweight solution where I can download, manage and run local models. Really not a fan of boarding the enshittification train ride with them.
Always had a bad feeling when they didn't give ggerganov/llama.cpp their deserved credit for making Ollama possible in the first place, if it were a true OSS project they would have, but now makes more sense through the lens of a VC-funded project looking to grab as much marketshare as possible to avoid raising awareness for alternatives in OSS projects they depend on.
Together with their new closed-source UI [1] it's time for me to switch back to llama.cpp's cli/server.
> Repackaging existing software while literally adding no useful functionality was always their gig.
Developers continue to be blind to usability and UI/UX. Ollama lets you just install it, just install models, and go. The only other thing really like that is LM-Studio.
It's not surprising that the people behind it are Docker people. Yes you can do everything Docker does with Linux kernel and shell commands, but do you want to?
Making software usable is often many orders of magnitude more work than making software work.
Can it easily run as a server process in the background? To me, not having to load the LLM into memory for every single interaction is a big win of Ollama.
I'm not sure which package we use that is triggering this. My guess is llama.cpp based on what I see on social? Ollama has long shifted to using our own engine. We do use llama.cpp for legacy and backwards compatibility. I want to be clear it's not a knock on the llama.cpp project either.
There are certain features we want to build into Ollama, and we want to be opinionated on the experience we want to build.
Have you supported our past gigs before? Why not be more happy and optimistic in seeing everyone build their dreams (success or not).
If you go build a project of your dreams, I'd be supportive of it too.
Why does everything AI-related have to be $20? Why can't there be tiers? OpenAI setting the standard of $20/m for every AI application is one of the worst things to ever happen.
I should have specified less expensive tiers (below the $20 standard). A tier <= $10 would be great. Anything over $10 for casual use seems excessive (or at least from my perspective)
My guess is that’s the lowest price point that provides a modicum of profitability — LLMs are quite expensive to run, and even more so for providers like Ollama, which are entering the market and don’t have idle capacity.
Claude has $20, $100 and $200, ChatGPT $20, and $200, Google has $20 and $250. Those all have free tiers as well, and metered APIs. Grok has $30 and $300 it looks like, the list probably goes on and on.
More than one year in and Ollama still doesn't support Vulkan inference. Vulkan is essential for consumer hardware. Ollama is a failed project at this point: https://news.ycombinator.com/item?id=42886680
There's an open pull request https://github.com/ollama/ollama/pull/9650 but it needs to be forward ported/rebased to the current version before the maintainers can even consider merging it.
Also realistically, Vulkan Compute support mostly helps iGPU's and older/lower-end dGPU's, which can only bring a modest performance speed up in the compute-bound preprocessing phase (because modern CPU inference wins in the text-generation phase due to better memory bandwidth). There are exceptions such as modern Intel dGPU's or perhaps Macs running Asahi where Vulkan Compute can be more broadly useful, but these are also quite rare.
I’m not throwing the towel on Ollama yet. They do need dollars to operate, but still provide excellent software for running models locally and without paying them a dime.
Ollama at its core will always be open. Not all users have the computer to run models locally, and it is only fair if we provide GPUs that cost us money and let the users who optionally want it to pay for it.
I think it’s the logical move to ensure Ollama can continue to fund development. I think you will probably end up having to add more tiers or some way for users to buy more credits/gpu time. See anthropic’s recent move with Claude code due to the usage of a number of 24/7 users.
- Speed
- Cost
- Reliability
- Feature Parity (eg: context caching)
- Performance (What quant level is being used...really?)
- Host region/data privacy guarantees
- LTS
And that's not even including the decision of what model you want to use!
Realistically if you want to use an OSS model instead of the big 3, you're faced with evalutating models/providers across all these axes, which can require a fair amount of expertise to discern. You may even have to write your own custom evaluations. Meanwhile Anthropic/OAI/Google "just work" and you get what it says on the tin, to the best of their ability. Even if they're more expensive (and they're not that much more expensive), you are basically paying for the priviledge of "we'll handle everything for you".
I think until providers start standardizing OSS offerings, we're going to continue to exist in this in-between world where OSS models theoretically are at performance parity with closed source, but in practice aren't really even in the running for serious large scale deployments.
[1] https://californiarecorder.com/sam-altman-requires-ai-privil...
> The order, embedded under and issued on Might 13, 2025, by U.S. Justice of the Peace Decide Ona T. Wang
Is this some meme where “may” is being replaced with “might”, or some word substitution gone awry? I don’t get it.
Sure, llama.cpp is the real thing, ollama is a wrapper... I would never want to use something like ollama in a production setting. But if I want to quickly get someone less technical up to speed to develop an LLM-enabled system and run qwen or w/e locally, well then its pretty nice that they have a GUI and a .dmg to install.
Since the new multimodal engine, Ollama has moved off of llama.cpp as a wrapper. We do continue to use the GGML library, and ask hardware partners to help optimize it.
Ollama might look like a toy and what looks trivial to build. I can say, to keep its simplicity, we go through a deep amount of struggles to make it work with the experience we want.
Simplicity is often overlooked, but we want to build the world we want to see.
Where can I learn more about this? llama.cpp is an inference application built using the ggml library. Does this mean, Ollama now has it's own code for what llama.cpp does?
I knew a startup that deployed ollama on a customers premises and when I asked them why, they had absolutely no good reason. Likely they did it because it was easy. That's not the "easy to use" case you want to solve for.
Why does this matter? For this specific release, we benchmarked against OpenAI’s reference implementation to make sure Ollama is on par. We also spent a significant amount of time getting harmony implemented the way intended.
I know vLLM also worked hard to implement against the reference and have shared their benchmarks publicly.
We can obviously disagree with their priorities, their roadmap, the fact that the client isn't FOSS (I wish it was!), etc but no one can say that ollama doesn't work. It works. And like mchiang said above: its dead simple, on purpose.
(any differences are small enough that they either shouldn't cause the human much work or can very easily be delegated to AI)
Then you want to swap models on the fly. llama-swap you say? You now get to learn a new custom yaml based config file syntax that does basically nothing that the Ollama model file already does so that you can ultimately... have the same experience as Ollama but now you've lost hours just to get back to square one.
Then you need it to start and be ready with the system reboot? Great, now you get to write some systemd services, move stuff into system-level folders, create some groups and users and poof, there goes another hour of your time.
GGML library is llama.cpp. They are one and the same.
Ollama made sense when llama.cpp was hard to use. Ollama does not have value preposition anymore.
The models are implemented by Ollama https://github.com/ollama/ollama/tree/main/model/models
I can say as a fact, for the gpt-oss model, we also implemented our own MXFP4 kernel. Benchmarked against the reference implementations to make sure Ollama is on par. We implemented harmony and tested it. This should significantly impact tool calling capability.
Im not sure if im feeding here. We really love what we do, and I hope it shows in our product, in Ollama’s design and in our voice to our community.
You don’t have to like Ollama. That’s subjective to your taste. As a maintainer, I certainly hope to have you as a user one day. If we don’t meet your needs and you want to use an alternative project, that’s totally cool too. It’s the power of having a choice.
Is there a schedule for adding additional models to Turbo mode plan, in addition to gpt-oss 20/120b? I wanted to try your $20/month Turbo plan, but I would like to be able to experiment with a few other large models.
Nope…
We benchmarked vLLM and Ollama on both startup time and tokens per seconds. Ollama comes at the top. We hope to be able to publish these results soon.
If you can't get access to "real" datacenter GPUs for any reason and essentially do desktop, clientside deploys, it's your best bet.
It's not a common scenario, but a desktop with a 4090 or two is all you can get in some organizations.
I'm also interested to see if that small minority of people are willing to pay for a service like this.
The solution is having proper governance for OSS projects that matter with independent organizations made up of developers, companies, and users taking care of the governance. A lot of projects that have that have last for decades and will likely survive for decades more.
And part of that solution is to also steer clear of projects without that. I've been burned a couple of times now getting stuck with OSS components where the license was changed and the companies behind it had their little IPOs and started serving share holders instead of users (elastic, redis, mongo, etc). I only briefly used Mongo and I got a whiff of where things were going and just cut loose from it. With Elastic the license shenenigans started shortly after their IPO and things have been very disruptive to the community (with half using Opensearch now). With Redis I planned the switch to Valkey the second it was announced. Clear cut case of cutting loose. Valkey looks like it has proper governance. Redis never had that.
Ollama seems relatively OK by this benchmark. The software (ollama server) is MIT licensed and there appears to be no contributor license agreement in place. But it's a small group of people that do most of the coding and they all work for the same vc funded company behind ollama. That's not proper governance. They could fail. They could relicense. They could decide that they don't like open source after all. Etc. Worth considering before you bet your company on making this a foundational piece of your tech stack.
I looked up the Ollama trademark and was surprised to see it's a Canadian company.
For Draw Things provided "Cloud Compute", we don't retain any data too (everything is done in RAM per request). But that is still unsatisfactory personally. We will soon add "privacy pass" support, but still not to the satisfactory. Transparency log that can be attested on the hardware would be nice (since we run our open-source gRPCServerCLI too), but I just don't know where to start.
[full disclosure I am working on something with actual privacy guarantees for LLM calls that does use a transparency log, etc.]
Edit: emailed the address on the site in your profile, got an inbox does not exist error.
This allows you to try out some open models and better assess if you could buy a dgx box or Mac Studio with a lot of unified memory and build out what you want to do locally without actually investing in very expensive hardware.
Certain applications require good privacy control and on-prem and local are something certain financial/medical/law developers want. This allows you to build something and test it on non-private data and then drop in real local hardware later in the process.
I feel like they're competing against Hugging Face or even Colaboratory then if this is the case.
And for cases that require strict privacy control, I don't think I'd run it on emergent models or if I really have to, I would prefer doing so on an existing cloud setup already that has the necessary trust / compliance barriers addressed. (does Ollama Turbo even have their Trust center up?)
I can see its potential once it gets rolling, since there's a lot of ollama installations out there.
For the users who want GPUs, which cost us money, we will charge money for it. Completely optional.
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
I moved away from ollama in favor of llama-server a couple of months ago and never missed anything, since I'm still using the same UI.
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)
> We do use GGML
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."
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.
thanks for the feedback address :)
That is interesting, did Ollama develop its own proprietary inference engine or did you move to something else?
Any specific reason why you moved away from llama.cpp?
Assuming you have llama-server installed, you can download + run a hugging face model with something like
And access http://localhost:8080Is it bad to fairly charge money for selling GPUs that cost us money too, and use that money to grow the core open-source project?
At one point, it just has to be reasonable. I'd like to believe by having a conscientious, we can create something great.
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.
They have very detailed quick start docs on it: https://docs.openwebui.com/getting-started/quick-start/start...
Why? If the tool works then use it. They’re not forcing you to use the cloud.
Its a tale we seen played out many times. Redis is the most recent example.
For one of the top local open model inference engines of choice - only supporting OSS out of the gate feels like an angle to just ride the hype knowing OSS is announced today "oh OSS came out and you can use Ollama Turbo to use it"
The subscription based pricing is really interesting. Other players offer this but not for API type services. I always imagine that there will be a real pricing war with LLMs with time / as capabilities mature, and going monthly pricing on API services is possibly a symptom of that
What does this mean for the local inference engine? Does Ollama have enough resources to maintain both?
I pay $20 to Anthropic, so I don’t think I’d get enough use out of this for the $20 fee. But being able to spin up any of these models and use as needed (and compare) seems extremely useful to me.
I hope this works out well for the team.
Usage-based pricing would put them in competition with established services like deepinfra.com, novita.ai, and ultimately openrouter.ai. They would go in with more name-recognition, but the established competition is already very competitive on pricing
Agreed, though there are already several providers of these new OpenAI models available, so I'm not sure what ollama's value add is there (there are plenty of good chat/code/etc interfaces available if you are bringing your own API keys).
He (almost) single-handedly brought LLMs to the masses.
With the latest news of some AI engineers' compensation reaching up to a billion dollars, feels a bit unfair that Georgi is not getting a much larger slice of the pie.
Now I am going to go and write a wrapper around llamacpp, that is only open source, truly local.
How can I trust ollama to not to sell my data.
You don't need to use Turbo mode; it's just there for people who don't have capable enough GPUs.
> ggml.ai is a company founded by Georgi Gerganov to support the development of ggml. Nat Friedman and Daniel Gross provided the pre-seed funding.
I think he's happy doing his own thing.
But then, if someone came in with a billion ... who wouldn't give it a thought?
$50M, now thats just perfect. you're retired, nor burdened with a huge responsibility
I guess their target audience values convenience and easy of use above all else so that could play well there maybe.
Doesn't look that much better than a ChatGPT Plus subscription.
ps: looking for most economic one to play around with as long as it a decent enough experience (minimal learning curve). buy, happy to pay too
https://github.com/ollama/ollama/issues/5245
If any of the major inference engines - vLLM, Sglang, llama.cpp - incorporated api driven model switching, automatic model unload after idle and automatic CPU layer offloading to avoid OOM it would avoid the need for ollama.
However the approach to model swapping is not 'ollama compatible' which means all the OSS tools supporting 'ollama' Ex Openwebui, Openhands, Bolt.diy, n8n, flowise, browser-use etc.. aren't able to take advantage of this particularly useful capability as best I can tell.
There is obviously some connection to Llama (the original models giving rise to llama.cpp which Ollama was built on) but the companies have no affiliation.
> OpenAI and Ollama partner to launch gpt-oss
If I use local/OSS models it's specifically to avoid running in a country with no data protection laws. It's a big close miss here.
All things considered though, Europe is getting confusing. They have GDPR but now pushing to backdoor encryption within the EU? [1]
At least there isn't a strong movement in the US trying to outlaw E2E encryption.
[1] https://www.eff.org/deeplinks/2025/06/eus-encryption-roadmap...
Which brings up the point are truly private LLMs possible? Where the input I provide is only meaningful to me, but the LLM can still transform it without gaining any contextual value out of it? Without sharing a key? If this can be done, can it be done performantly?
Yes, there is gonna be a new discussion for it on October 15, but I've already seen section of governments being against their own government position on the bill (Swedish Military for example).
I guarantee that nobody cares about or will be surveilling your private AI use unless you're doing other things that warrant surveillance.
The reason big providers suck, as OpenAI is so nicely demonstrating for us, is that they retain everything, the user is the product, and court cases, other situations can unmask and expose everything you do on a platform to third parties. This country seriously needs a digital bill of rights.
The biggest game in town has been managing platforms that give owners an information advantage. But at least the world generally trusts the USA to abide by laws and user agreements, which is why, to my mind, the USA retains the near monopoly on information platforms.
I personally wouldn’t trust a UK platform for example, being a Brit native. The top echelon talent pool is so small and incestuous I don’t believe I would experience a fair playing field if a business of mine passed a certain size of national reach/importance.
EDIT: from ChatGPT, new money entrepreneurs with no inheritence/political ties by economic region, USA ~63%, UK/HongKong/Singapore ~45%, Emerging Markets ~35%, EU ~22%, Russia ~10%
In a universe where everything you say can be taken out of context, things like OpenAi will be a data leak nightmare.
Need this soon:
https://arxiv.org/abs/2410.02486
> Turbo is a new way to run open models using datacenter-grade hardware.
What? Why not just say that it is a cloud-based service for running models? Why this language?
Thankfully, this may just leave more room for other open source local inference engines.
The Ollama app using the signed-in-only web search tool is really pretty good.
There are areas we will make money, and I wholly believe if we follow our conscious we can create something amazing for the world while making sure we can keep it fueled to keep it going for the long term.
Some of the ideas in Turbo mode (completely optional) is to serve the users who want a faster GPU, and adding in additional capabilities like web search. We loved the experience so much that we decided to give web search to non-paid users too. (Again, it's fully optional). Now to prevent abuse and make sure our costs don't go out of hand, we require login.
Can't we all just work together and create a better world? Or does it have to be so zero sum?
For Turbo mode I understand the need for paying but the main poing of running a local model with web search is browsing from my computer without using any LLM provider. Also I want to get rid of the latency to US servers from Europe.
If ollama can't do it, maybe a fork.
Wait until it makes significant amounts of money. Suddenly the priorities will be different.
I don’t begrudge them wanting to make some money off it though.
This isn't Anaconda, they didn't do a bait and switch to screw their core users. It isn't sinful for devs to try and earn a living.
If you earn a living using something someone else built, and expect them not to earn a living, your paycheck has a limited lifetime.
“Someone” in this context could be a person, a team, or a corporate entity. Free may be temporary.
if i could have consistent and seamless local-cloud dev that would be a nice win. everyone has to write things 3x over these days depending on your garden of choice, even with langchain/llamaindex
It was always just a wrapper around the real well designed OSS, llama.cpp. Ollama even messes up the names of models by calling distilled models the name of the actual one, such as DeepSeek.
Ollama's engineers created Docker Desktop, and you can see how that turned out, so I don't have much faith in them to continue to stay open given what a rugpull Docker Desktop became.
Always had a bad feeling when they didn't give ggerganov/llama.cpp their deserved credit for making Ollama possible in the first place, if it were a true OSS project they would have, but now makes more sense through the lens of a VC-funded project looking to grab as much marketshare as possible to avoid raising awareness for alternatives in OSS projects they depend on.
Together with their new closed-source UI [1] it's time for me to switch back to llama.cpp's cli/server.
[1] https://www.reddit.com/r/LocalLLaMA/comments/1meeyee/ollamas...
All companies that raise outside investment follow this route.
No exceptions.
And yes this is how ollama will fall due to enshittification, for lack of a better word.
https://news.ycombinator.com/newsguidelines.html
Developers continue to be blind to usability and UI/UX. Ollama lets you just install it, just install models, and go. The only other thing really like that is LM-Studio.
It's not surprising that the people behind it are Docker people. Yes you can do everything Docker does with Linux kernel and shell commands, but do you want to?
Making software usable is often many orders of magnitude more work than making software work.
So does the original llama.cpp. And you won't have to deal with mislabeled models and insane defaults out of the box.
No inference engine does all of:
- Model switching
- Unload after idle
- Dynamic layer offload to CPU to avoid OOM
I'm not sure which package we use that is triggering this. My guess is llama.cpp based on what I see on social? Ollama has long shifted to using our own engine. We do use llama.cpp for legacy and backwards compatibility. I want to be clear it's not a knock on the llama.cpp project either.
There are certain features we want to build into Ollama, and we want to be opinionated on the experience we want to build.
Have you supported our past gigs before? Why not be more happy and optimistic in seeing everyone build their dreams (success or not).
If you go build a project of your dreams, I'd be supportive of it too.
Docker Desktop? One of the most memorable private equity rugpulls in developer tooling?
Fool me once shame on you, fool me twice shame on me
llama.cpp isn't (just) a C++ library/codebase -- it's a CLI application, server application (llama-server), etc.
https://www.anthropic.com/pricing - $0 / $17 (if billed annually) / $20 (if billed monthly) / $100 / $25 (team) / custom enterprise pricing / on-demand API pricing
Sounds like tiers to me.
Also realistically, Vulkan Compute support mostly helps iGPU's and older/lower-end dGPU's, which can only bring a modest performance speed up in the compute-bound preprocessing phase (because modern CPU inference wins in the text-generation phase due to better memory bandwidth). There are exceptions such as modern Intel dGPU's or perhaps Macs running Asahi where Vulkan Compute can be more broadly useful, but these are also quite rare.
Is it because they developed s new ollama which isn't open and which doesn't use llama.cpp?
It is completely compromised, especially if it is an AI company.
How do you think ollama was able to provide the open source AI models to everyone for free?
I am pretty sure ollama was losing money on every pull of those images from their infrastructure.
Those that are now angry at ollama charging money or not focusing on privacy should have been angry when they raised money from investors.
If you want to see where the actual developers do the actual hard work, go use llama.cpp instead.