Ollama

Where does your data go?

Nowhere — that's the entire product. But the honest answer has three layers:

• Locally: nothing leaves your machine. Ollama runs open-weight models on your own hardware. No account, no login, no telemetry requirement. What you type stays in RAM and your disk.
• Cloud tier (optional, separate): Ollama sells cloud-hosted inference — hosted in the US and Europe (Singapore for some models), with a published zero data retention commitment: no prompt logging, no training on your data, per-request cost visible in your account.
• The catch nobody advertises: defaults are insecure. The local API has no authentication, and the documented OLLAMA_HOST=0.0.0.0 setup — widely used in practice — exposes it to the network. Security researchers counted ~300,000 Ollama servers publicly reachable on the internet.

What happens when it's wrong?

A local model's failure mode is the model, not the vendor — and open models are measurably weaker than the closed flagships:

• Quality is hardware-capped. A 7–8B model (fits 16GB RAM/6GB VRAM) is genuinely useful for drafting and summarizing but nowhere near ChatGPT or Claude flagship level on reasoning. A 70B model rivals them — and needs 64GB RAM or 48GB+ VRAM. Pick the biggest model your machine fits, and expect a gap.
• Documented security failures — the section every local-first pitch skips. CVE-2026-7482 ("Bleeding Llama," CVSS 9.1, patched Feb 2026): an unauthenticated attacker could leak other users' conversation history, system prompts, and environment variables — API keys included — from any reachable Ollama server in three API calls.
• Worse for Windows users: the updater is still unpatched. CVE-2026-42248/42249 (disclosed by CERT Polska after 90 days of maintainer silence): the Windows auto-updater doesn't verify update signatures and writes installers from HTTP headers — a chain that drops a persistent executable on every login.
• What the failure teaches: the tool's security posture depends on patch discipline and network hygiene that a small-business owner may not have.

What's it actually bad at?

• The app has closed much of the gap, but depth still needs the ecosystem. The desktop app now offers chat with file drag-and-drop and even connects Claude Desktop to your local models; Open WebUI remains the answer for advanced workflows. Setup is easy; going deep is homework.
• Hardware is the real price. 16GB RAM for a useful model, 64GB for a flagship-class one, an SSD strongly recommended. The software is free; the machine that runs it well is not.
• Model quality varies wildly — open models span from excellent to toy, and there's no vetting layer between "downloads on ollama.com" and "worth trusting."
• Windows auto-update is currently unsafe — the one platform where the default install needs manual intervention.
• The maintainer-response record is mixed: CERT Polska reported the Windows flaws in January 2026; 90 days passed with no engagement before they published. Open-source speed cuts both ways.

What does it cost really?

The traps: the pricing page now sells cloud first — easy to mistake Ollama for a paid product when local is the headline. Cloud concurrency is limited (1 free / 3 Pro / 10 Max), and peak pricing applies to cloud models at midday UTC. The "Free" cloud tier is a trial of starter models, not the local product.

What does it replace — and what do you still do?

It replaces the subscription relationship: a capable model that answers to nobody's server, no usage anxiety, no policy changes, works offline. For regulated data — customer records, patient info, legal documents — it's the only approach where the data-handling answer is "it never leaves the building." You still own: hardware maintenance, model selection, updating the software, and the network security the defaults don't provide.

What the ecosystem offers

• The desktop app closed the gap: a proper chat interface with file drag-and-drop (text, PDFs, images), a setup walkthrough for first launch, and faster responses — model loading got roughly twice as quick in a recent release.
• Claude Desktop can use it: Anthropic's desktop app now connects to Ollama as a gateway, meaning you can use your own local models inside Claude's interface — local and hosted tools, working together.
• Agent launchers: ollama launch runs coding agents and DeepSeek's agent harness against your local models; the OpenAI-compatible API gained web search; Mac performance improved through a new Apple-chip engine.
• Release cadence: three updates in the review week — fixes landing fast, which is the open-source trade: quick patches, but you're the one applying them.
• Still standing: the Windows updater security flaws remain unpatched as of this re-review.

If I stop paying, what do I lose?

• Nothing — there's nothing to cancel. Local Ollama has no subscription, no account, no server-side state. Delete the app and the models; your data was always on your machine.
• The cloud tier exits cleanly: cancel and keep using local; zero data retention means there's no history to lose either way.
• The lock-in reality is model formats, not the vendor: models run anywhere llama.cpp-derived runtimes run — switching to alternatives (LM Studio, vLLM) is friction, not loss. The MIT license and open weights mean no exit cliff exists.

Who is this NOT for?

Anyone who won't maintain software — unpatched Ollama is a data leak, and the Windows updater is currently unsafe. Businesses without decent hardware — a model that fits in 8GB RAM is a demo, not an assistant. And anyone who expected "local" to mean "zero homework": the security defaults are the homework.