Ollama
Version in force 0.40.102026-10-23Tree audited tag v0.40.2 (commit b061384d90ff455462bc32745dd0de479de717a3)Licence MIT
Ollama 0.40.10 runs models on your Windows, Mac or Linux machine, but local does not mean silent. No third-party analytics; yet out of the box it makes up to 4 calls to ollama.com, including the desktop app’s hourly update check, which cannot be switched off. Its code is public and MIT-licensed, so we checked these calls by reading it.
A local model server with a desktop installer, its own catalogue and one command that downloads and runs. It is the closest thing there is to a black box that just works.
Steps to the first token
Measured by Local AI Scope
Hosts it contacts out of the box
Vendor declared
Licence
Measured by Local AI Scope
Videos: Ollama 4
The best walk-throughs we found. Press play and it opens right here.
The videos live on YouTube, and nothing loads from there until you press play (no-cookie mode). The thumbnail is served by us.
How we checked this
Everything this sheet says about the network comes from reading the published source code and the official documentation on the date shown — not from watching the traffic. We have not yet sat the program in front of a network analyser, so this is static analysis, not observation. That capture is still on our list; the day we have it, these cells move up a level of evidence. Until then, no pretending we sniffed the wire: we have read the recipe, not tasted the dish.
What leaves your machine
Before a local runtime gets anywhere near your documents, we ask it four questions: what it sends on its own, who else it calls, which address it listens on, and whether you can check any of that yourself. Think of it as a job interview.
Telemetry out of the box
No third-party analytics — but it is not silent either. Verified by us in the primary source
There is not a single analytics library anywhere in the published tree: no telemetry, analytics, posthog, mixpanel, sentry or datadog outside test files. That is a point in its favour and it is checkable. What there is instead are automatic calls to ollama.com, which is a different thing and is covered in the next section.
«We may collect limited device and usage metadata (such as app version and request counts).»
How to switch it off: There is no telemetry switch: none of the 27 OLLAMA_* environment variables Ollama reads is one, because there is no analytics SDK to switch off. The lever that does exist, OLLAMA_NO_CLOUD=1, cuts the cloud features and the remote catalogue — not the update check, and not the account check the desktop app makes when it opens.
Source · consulted 2026-10-10
Other calls home
| What | Where to | When | Can you avoid it | Evidence |
|---|---|---|---|---|
| Model catalogue and downloads | registry.ollama.ai |
When you pull or run a model that is not on disk. | No — that is the feature | verified by us |
| Update check, signed with a key stored on your machine | ollama.com/api/update |
Every hour while the desktop app is open. On macOS it also carries a persistent device identifier. | No. See below | verified by us |
| Remote catalogue of recommended models, signed with your installation key since 0.33.0 | ollama.com/api/experimental/model-recommendations |
At start-up and then every 4 hours, started by `ollama serve` itself — no desktop app and no account needed. | Yes: OLLAMA_NO_CLOUD=1 | verified by us |
| Cloud inference and web search | ollama.com |
Only with :cloud models or web search, and only after `ollama signin`. | Yes: OLLAMA_NO_CLOUD=1 | vendor declared |
| Preloading the list of cloud models, signed | ollama.com/api/tags · ollama.com/api/show |
Once each time `ollama serve` starts — Linux included, no app and no account — and when you open a :cloud model. | Yes: OLLAMA_NO_CLOUD=1 | verified by us |
| Account check when the app opens, signed | ollama.com/api/me |
When the macOS or Windows desktop app opens, and again when its window asks for the user (at most every 5 minutes) — with or without an account. | No | verified by us |
| Catalogue for Claude Desktop and the ChatGPT app, signed | ollama.com/api/experimental/model-recommendations?app=… · ollama.com/api/tags |
macOS only, and only if you connect Claude Desktop or the ChatGPT app to Ollama from the app. | Yes: do not connect that integration | verified by us |
«The original macOS app used to use the device ID to check for updates so include it if present. […] Auto-update disabled – don’t download, just log.»
Source · consulted 2026-10-10
Which address it listens on
| Setting | Out of the box |
|---|---|
| Listens on | 127.0.0.1:11434 |
| Authentication out of the box | none |
| CORS out of the box | Permissive: localhost on any port, plus app://, file://, tauri://, vscode-webview:// and vscode-file://; browser extensions only if added to OLLAMA_ORIGINS |
It listens on localhost only, which is the sensible default. The loose end is CORS: without configuring anything it accepts requests from any page served on localhost at any port, plus app://, file://, tauri://, vscode-webview:// and vscode-file://; browser extensions get a 403 until you add them to OLLAMA_ORIGINS. And the local API has no authentication of any kind — OLLAMA_AUTH does not protect the server, it only makes the client sign its requests. Since version 0.40.1 (7 October 2026), the local server also answers GET /api/balance and GET /api/usage by forwarding them to ollama.com signed with the machine’s own key, so whoever reaches the port, with no key, reads the cloud balance and usage of the account this machine is signed in to. OLLAMA_NO_CLOUD=1 closes both with a 403 without leaving the machine. Verified by us in the primary source
Source · consulted 2026-10-10
Source · consulted 2026-10-10
Can this be audited?
Yes. The code is public and MIT, so every statement above about the network was checked by reading it rather than believing it. That is the difference between a fact and a promise, and it is why this row of the table can be trusted at all. Verified by us in the primary source
Source · consulted 2026-10-11
Privacy policy
Ollama Inc., policy dated March 2026. It says the program runs locally and that they do not see your prompts, and at the same time that they may collect limited device and usage metadata — including request counts, which implies counting per installation. vendor declared Source
What it runs, and where
Platforms, accelerators and model formats: the three walls you can hit on that first, eager evening. Better to find them here than at midnight with a model half downloaded.
Where it runs
| Platform | Support | Evidence |
|---|---|---|
| Windows | Yes — Windows 10 22H2 or newer, x64 and ARM64. Installing does not require administrator rights. | vendor declared |
| macOS Apple Silicon | Yes — macOS Sonoma (14) or newer, using both CPU and GPU. Since 0.40.0, models that have an MLX version run on MLX by default. | vendor declared |
| macOS Intel | Yes, but CPU only according to Ollama: the macOS build is universal, so the binary runs. | vendor declared |
| Linux | Yes — tarballs for amd64 and arm64, a separate ROCm package for AMD cards and JetPack packages for Jetson. No desktop app on Linux. | verified by us |
What it accelerates
| Backend | Status | Evidence |
|---|---|---|
| CUDA | Yes — NVIDIA with compute capability 5.0+ and driver 550+. Cards from 5.0 to 6.2 need driver 570+. | vendor declared |
| ROCm | Yes, on Windows too — needs the ROCm v7 / HIP7 driver stack. | vendor declared |
| Metal | Yes on Apple Silicon: through llama.cpp, and since 0.40.0 through its MLX engine by default for the models that have an MLX version. No Metal on Intel Macs. | vendor declared |
| Vulkan | Yes, and it is on by default on Windows and Linux. Switch it off with OLLAMA_VULKAN=0. | verified by us |
| CPU | Yes, with libraries per vectorisation level. | verified by us |
| NPU | No. No Ryzen AI NPU, no Apple Neural Engine, no QNN, DirectML or OpenVINO. | verified by us |
Model formats it runs
GGUF and safetensors. Catalogue models come in its own layered packaging, and since 0.40.0 a model can come in two versions, GGUF for llama.cpp and safetensors for MLX: an Apple Silicon Mac downloads the MLX one when there is one, everything else the GGUF. To bring your own, it takes a safetensors directory or a .gguf file. vendor declared
Source · consulted 2026-10-07
Installing it, and living with it
The everyday part: how many steps to the first token, what it leaves on your disk, under which licence, and at what pace it changes under your feet once it is installed.
Steps to the first token
2 steps. Install it and run `ollama run <model>`: one command downloads the model and answers. verified by us
Step by step: install it, check it is running, and switch off what can be switched off
If something fails: 9 errors, the exact message each one shows and the fix
What it writes to your disk
| What | Where | Evidence |
|---|---|---|
| Models | ~/.ollama/models — configurable with OLLAMA_MODELS; on Linux run as a service, /usr/share/ollama/.ollama/models |
verified by us |
| The key that identifies your installation to ollama.com | ~/.ollama/id_ed25519 + .pub |
verified by us |
| Chat history of the desktop app — an unencrypted SQLite database | macOS ~/Library/Application Support/Ollama/db.sqlite · Windows %LOCALAPPDATA%\Ollama\db.sqlite · Linux ~/.ollama/db.sqlite |
verified by us |
| Command-line history | ~/.ollama/history — off with OLLAMA_NOHISTORY |
verified by us |
| Cached remote catalogue | ~/.ollama/cache/model-recommendations.json |
verified by us |
| Logs and configuration | ~/.ollama/logs/{app,server,upgrade}.log · ~/.ollama/config.json · ~/.ollama/server.json |
vendor declared |
Licence
MIT. Free to use, modify and redistribute, at work included. verified by us
Source · consulted 2026-10-11
Version and release pace
Very fast: two or three stable releases a week in bursts, and roughly a minor version bump per month. 104 releases in twelve months, 99 of them stable. verified by us
Source · consulted 2026-08-25
What to know before you install it
- Local does not mean silent. The bare server — no desktop app, no account — calls ollama.com at start-up and then every 4 hours for its remote catalogue, and since version 0.33.0 that request is signed with your installation key. That one you can cut with OLLAMA_NO_CLOUD=1.
- The desktop app checks for updates every hour, the request is signed with a key stored on your machine, and that check cannot be turned off. The auto-update setting only stops the download: the loop calls out first and reads the setting afterwards.
- The server listens on 127.0.0.1 only, but it accepts requests from any page served on localhost and from desktop apps that load their pages as file://. A local web app served on localhost can talk to your model without asking you. Browser extensions are the exception: they get a 403 until you add them to OLLAMA_ORIGINS.
- The local API has no authentication whatsoever. OLLAMA_AUTH does not protect the server; it only makes the client sign.
- Models with the :cloud suffix do not run on your machine — they are sent to Ollama’s servers. Same command line, completely different privacy.
- When it opens, the desktop app asks ollama.com about your account with a signed request, even if you have no account. OLLAMA_NO_CLOUD does not stop it.