LOCAL AI TOOLS · SEPTEMBER 2026
Should You Stop Using Ollama? I Checked Every Claim
A viral essay says switch immediately. Three of its five accusations were already fixed — and one “free” alternative it recommends costs $149 a year if you use it for work.
I read the essay, then checked each claim against the actual repositories, licenses and terms of service. Here’s what survived.
The short version
Probably not. The essay’s criticism of how Ollama’s team responds to problems is fair and still stands. But the technical complaints have mostly been fixed since it was written, the famous “1.8× faster” number can’t be traced to any real benchmark, and Ollama is still free and unlimited for local use.
If you run AI models on your own computer, you probably use Ollama. It’s the one that just works — you type one command, and a model runs. That’s why it won.
So when an essay titled “Friends Don’t Let Friends Use Ollama” started circulating again this month, it landed hard. Hundreds of upvotes. A long comment thread. And a lot of people quietly wondering whether they’d bet on the wrong tool.
I went and checked it, claim by claim, against the actual source code and license files. The answer turned out to be more interesting than either “it’s fine” or “switch now.”
First: the essay is older than it looks
This matters more than anything else, so I’m putting it first.
The essay was published in April 2026. What’s circulating now is a re-share. Five months in open-source time is a long while, and several of its central complaints describe a version of Ollama that no longer exists.
That’s not the author’s fault — it was accurate when written. But if you’re reading it today as a description of the current state of things, you’re reading a snapshot, not a photograph of now.
Three complaints that have since been fixed
“It doesn’t credit llama.cpp”
Ollama’s speed comes from llama.cpp, an open-source project by Georgi Gerganov. For a long stretch, Ollama barely acknowledged it — and the essay documents that history well, including a request that sat unanswered for over a year.
Today the README has a section headed “Supported backends” that names llama.cpp and credits Gerganov by name, sitting near the top of the file rather than buried at the bottom. The historical criticism is fair. The present-tense version isn’t.
“The desktop app is closed-source”
It isn’t. The repository is MIT-licensed and the desktop app folder is in it. The essay itself notes this was resolved in late 2025 — but the section heading still says “The Closed-Source App,” and that heading is what gets screenshotted.
“It abandoned llama.cpp for its own engine”
Ollama did build its own engine, and that caused real friction. But the repository now pins a specific upstream llama.cpp build again, alongside Apple’s MLX for Mac chips. The honest current criticism is narrower and more useful: the pinned version runs roughly seven weeks behind upstream, so brand-new model architectures show up in llama.cpp before they show up in Ollama.
The complaint that actually still stands
Here’s where the essay is right, and where I’d push harder than it does.
A security flaw was reported to Ollama in December 2024 — a bug where a malicious model registry could trick Ollama into handing over your authentication token. A volunteer from the community wrote the fix and opened it in May 2025.
It was merged in January 2026. Thirteen months from report to fix, on someone else’s patch, with pings along the way that went unanswered. A maintainer’s comment when it finally landed: “Apologies for the delay here.”
Separately, a request for a simple “export my model” command sat open from 2023 until April 2026, when it was closed as “not planned.”
Neither of these will slow down your laptop. But if you’re deciding what to build a habit on, how a project responds when something is wrong is a reasonable thing to weigh. That criticism has not been fixed, and Ollama has never publicly answered it.
Where the “1.8× faster” number actually came from
You’ve seen this figure. llama.cpp is 1.8 times faster than Ollama — 161 tokens per second versus 89. It gets quoted everywhere.
I tried to find the benchmark. The earliest version I could locate is a 2024 marketing page for a chatbot company, in a sentence that begins “in a recent benchmark” and “reportedly.” No link. No hardware named. No model named — just “a quantized model.” No context length, no number of runs, no date for the benchmark it’s supposedly reporting.
A second figure in the same essay refers to a model called Qwen-3 Coder 32B. That model doesn’t exist. Qwen3-Coder shipped at other sizes entirely.
So what do the tests that publish their methods find? Between 2% and 15%, depending on hardware and model — and in one careful test, Ollama came out slightly ahead.
⚠️ The catch behind the big numbers
One researcher published a 2.4× result, then retracted it himself. He’d discovered Ollama had quietly split the model between his graphics card and system memory because it didn’t fit — so he was comparing a fast run against a crippled one. Cleaned up, the gap collapsed. That’s the shape of most dramatic benchmark claims here.
The “free” alternative that isn’t actually free
This is the finding I did not expect, and it’s the one most likely to cost a reader money.
The essay recommends Msty among its alternatives to a venture-backed tool with paid tiers. Msty’s own terms of service say its free tier is for “private, non-commercial purposes” and that using it “for the exercise of your trade or profession for which you are compensated” does not qualify.
In plain terms: if you use Msty for work you get paid for, you need a licence. That’s $149 per year, or $349 once.
Ollama has no such restriction. It’s MIT-licensed, and its own pricing page says local use is “always unlimited.” So the essay recommends, as an escape from commercialisation, a tool with a stricter commercial-use restriction than the one it’s telling you to leave.
Yes, there’s a paid cloud now. No, that didn’t take your free version away.
Ollama raised $88 million and launched paid tiers — $20, $100 and $500 a month. That part is true, and it’s a fair thing to feel uneasy about.
But for a non-technical reader, “pivoted to cloud” reads as “it’s not free anymore,” and that isn’t what happened. Running models on your own machine requires no account, no payment, and no cloud connection. You can switch cloud off entirely with one line in a settings file.
Worth knowing: Ollama does collect some usage and download metadata by default. Its documentation is explicit that prompts and responses stay local — “We don’t see your prompts or data when you run locally” — but “no content leaves” is not the same as “nothing leaves.” If that distinction matters to you, llama.cpp and Jan send nothing at all.
What you can realistically install
The essay names seven alternatives. Four of them are developer tools that a normal person cannot install unaided — one requires Docker, another is a proxy that needs a hand-written config file, and llama.cpp itself has no installer at all.
LM Studio — the easiest switch
A polished desktop app with a model browser and a built-in chat window. Free, including at work — the requirement to contact them for a commercial licence was removed in July 2025. It’s proprietary software rather than open source, which may or may not matter to you.
One catch that catches people out: Intel Macs are not supported. Apple Silicon only.
Jan — the open-source one that still feels like an app
Apache 2.0 licensed, no account required, genuinely pleasant to use. If you want something open that a normal person can install, this is the answer. Same catch: Apple Silicon only on Mac.
koboldcpp — for Windows with an NVIDIA card
A single downloadable file, no installation. Rougher around the edges, and you have to pick the right download from several options.
⚡ Download koboldcpp from GitHub only
The project’s own README carries a phishing warning: a lookalike domain is impersonating it. Get it from the releases page on its official GitHub repository and nowhere else.
If you do switch, your models aren’t trapped
The essay suggests Ollama’s storage format locks you in. It doesn’t, quite. The files are stored under scrambled names, but the contents are ordinary model files that other tools can open directly.
1. ollama list — see what you have.
2. ollama show <model> --modelfile — the first line prints the real file path.
3. Copy that file somewhere and rename it to end in .gguf. Copy — don’t create a shortcut. Ollama deletes these files during cleanup, and a shortcut would break.
4. Open it in LM Studio or Jan. Confirm it works before deleting anything.
Honestly though — for most common models, re-downloading is less hassle than hunting through folders, and you get a clearly-named file at the end of it.
⚡ Don’t delete the whole Ollama folder
That folder also holds a signing key that can’t be regenerated. If you’re clearing space, delete only the models folder inside it.
And the thing nobody mentions: anything you’ve pointed at Ollama will break. Every alternative listens on a different port, so browser front-ends, editor plugins and scripts all need reconfiguring. For most people this is the real cost of switching, not the download.
The honest limitations
I didn’t run my own benchmark. I traced other people’s and reported which ones published their methods. Someone with one machine and a weekend could settle this properly, and nobody has.
The speed gap is real, just small. llama.cpp does win most like-for-like tests. If you’re running models all day at scale, 10% compounds. For chatting with a model, you won’t feel it.
Ollama has never responded. Not on its blog, not in the comment threads, and not on the two open performance reports on its own tracker. I checked. Silence isn’t a rebuttal, and it isn’t reassuring either.
Versions move. Everything here was checked in September 2026 against current repositories and terms. Licences and pricing change — the Msty terms I quote were themselves updated this month.
Who should actually care
Worth switching if: you want a nicer interface than a terminal, or you want the newest model architectures the week they land. LM Studio or Jan, and you’ll be running in ten minutes.
Maybe not if: you already have things wired into Ollama. Rewiring every plugin and script to a new port costs more than the 10% you’d gain.
Definitely stay if: you’re on an Intel Mac, or a Windows machine with an AMD graphics card. LM Studio and Jan won’t run on Intel Macs at all, and Ollama handles AMD setup for you.
If you haven’t tried running a model locally at all yet, start with my step-by-step guide to running a free local LLM with Ollama and Qwen — nothing in this article changes that recommendation for a first install.
Frequently asked questions
Is Ollama still free?
Yes. It’s MIT-licensed and running models on your own computer is free and unlimited, with no account required. Ollama also sells paid cloud tiers starting at $20 a month, and its own pricing page states that local use is “always unlimited.” Cloud features can be disabled entirely.
Is llama.cpp really faster than Ollama?
Usually, but by less than you’ve heard. Tests that publish their methods find roughly 2% to 15% on comparable hardware, and at least one found Ollama slightly ahead. The widely-quoted “1.8× faster, 161 versus 89 tokens per second” figure traces back to an uncited 2024 marketing page.
What’s the easiest Ollama alternative for a beginner?
LM Studio or Jan. Both have proper installers, a model browser and a chat window. Jan is open source (Apache 2.0) and requires no account. Neither runs on Intel Macs — Apple Silicon only.
Will I have to download all my models again if I switch?
No. Ollama stores standard model files under scrambled names, and ollama show <model> --modelfile reveals the path. Copy the file, rename it to end in .gguf, and open it in your new app. Copy rather than link, and confirm it loads before deleting anything.
The bottom line
The essay’s argument about how Ollama treats the people around it — the upstream project, the security reporter, the volunteer who wrote the patch — is worth taking seriously, and it hasn’t been answered. That’s a real reason to keep an eye on it.
The argument that you should switch tonight is weaker than it sounds. Three of the technical complaints are fixed, the speed number doesn’t hold up, Ollama is still free for local use, and one of the recommended alternatives would charge you $149 a year for the exact use case most readers have.
Your one next step: if you’ve been curious about a friendlier interface anyway, install Jan and point it at a model you already have. It’s free, open source, takes ten minutes, and you can keep Ollama installed while you decide.
