Local AI vs cloud AI: who can see your chats with each

The Morse teamPublished October 25, 20268 min read

Local AI vs cloud AI comes down to who holds your words. A model running on your own computer has no provider to read, keep, train on, or hand over your chats, so local AI is more private against the company. It isn't more private against everything: your history sits on your disk, local apps still talk to the internet for updates and downloads, and the strongest models don't run on a laptop at all. A cloud model inside a verified secure enclave sits between the two.

We read the network, privacy, and licence pages of four local tools (Ollama, LM Studio, Jan, and llama.cpp), plus a model card, on October 20, 2026, and built the table below from them. All four come out well, and none of them sends what you type to a local model anywhere.

Is local AI more private than cloud AI?

Yes, against the provider. What the four apps' network and privacy pages list as leaving your computer is mostly update checks, model downloads, and metadata, plus, in Jan, analytics you have to switch on; your chats with a local model aren't on any of those lists (the table below has each one). Jan's privacy page says conversation history and usage logs "are stored locally and never leave your computer" (Jan).

The licences decide how far you have to trust that. Ollama and llama.cpp are MIT and Jan is Apache 2.0, so anyone can read the code that makes those network calls. LM Studio is proprietary (terms), so for it you go on its own documentation, or watch its traffic yourself. Those terms license the app "solely for Your personal and / or internal business purposes," and say third-party AI models you download come to you under "a license agreement between you and the third party provider."

A cloud chatbot can't offer the same, because the model has to read your message on someone else's hardware. Our guide to using AI privately covers that basic trade.

What do local AI apps send over the network?

Less than a cloud chatbot, but not nothing. Here is what each tool's own pages say, read on October 20, 2026.

None of this exposes what you type to a local model. It does tell a server that your copy of the app exists, roughly where you are, and which models you fetch. Jan's README calls it an app that "runs 100% offline on your computer," and also lists hosted providers you can connect, so "offline" covers the local models only. Ollama's cloud models run on Ollama's servers, and its FAQ says it processes those prompts "to provide the service but do not store or log that content and never train on it" (Ollama), a promise you can't check, like any cloud host's.

To run AI locally for privacy with as little network traffic as possible, download the model, then cut the app off. LM Studio says it "can operate entirely offline, just make sure to get some model files first." Ollama logs "Ollama cloud disabled: true" once you set OLLAMA_NO_CLOUD=1 or disable_ollama_cloud in ~/.ollama/server.json, and you lose its cloud models and web search with it (Ollama). llama.cpp's server takes --offline, which "forces use of cache, prevents network access."

Can a local AI server leak your chats?

Yes, if you open it to the network. Each of these tools can serve an API to other programs: Jan at localhost:1337 (Jan), and Ollama on 127.0.0.1 port 11434 by default, changed with OLLAMA_HOST (Ollama). llama.cpp's server also listens on 127.0.0.1 by default and has no API key unless you set one with --api-key (llama.cpp).

Those defaults keep the server on your own machine, but people change them. On September 1, 2025, Cisco reported finding "1,139 vulnerable Ollama instances" reachable on the internet through Shodan, about 18.8% of them serving models, and blamed "default configurations, weak or absent authentication, and insufficient network isolation" (Cisco). Keep the bind address on 127.0.0.1 unless you know who else can reach it.

What hardware do you need to run AI locally?

Memory is the limit, and the model card tells you how much. OpenAI's card for gpt-oss-20b, an open-weight model with 21 billion parameters (3.6 billion active per token), says it can "run within 16GB of memory," while gpt-oss-120b fits "into a single 80GB GPU (like NVIDIA H100 or AMD MI300X)" (Hugging Face). Ollama's downloads of the same two models are 14 GB and 65 GB (Ollama). Both are Apache 2.0, so the model's licence is as loose as the app's.

For smaller models, Jan's README gives a Mac rule of thumb: "8GB RAM for 3B models, 16GB for 7B, 32GB for 13B" (Jan). On a PC, Ollama's GPU page sets the floor at Nvidia cards with "compute capability 5.0+ and driver version 550 and newer," with Apple GPUs covered through Metal and AMD through ROCm or Vulkan (Ollama). llama.cpp lists 17 backends, from CUDA and Metal to Vulkan and Snapdragon's Hexagon (llama.cpp).

So a 16 GB laptop runs a 20-billion-parameter model, and a 120-billion-parameter one needs a data-center GPU.

Who sees what: local vs cloud vs cloud with an enclave

Local or cloud has a different answer for each person you're worried about. This table is our reading of the pages above, as of October 2026. "Cloud" means a regular chatbot account; "enclave" means a model in a secure enclave that your device verifies before it encrypts the message (how that works).

Two rows favor local outright: the provider and the court's subpoena. Two favor neither: a device thief and an employer care about where history is stored, which is a separate choice from where the model runs. The network row barely changes. We cover how chats reach courtrooms in can AI chats be subpoenaed, and the limits of enclave hardware in what is a trusted execution environment.

When does cloud AI make more sense?

When you need a model your computer can't hold, or one that isn't released at all. gpt-oss-120b wants an 80 GB GPU, and the models behind ChatGPT, Claude, and Gemini can't be downloaded at any size. Cloud AI also brings web search, image generation, and long research tasks that a local setup has to assemble from parts. Then the privacy question is what that cloud learns about you, and a regular chatbot account is only one answer to it.

How Morse covers the cloud side

Every privacy cost of cloud AI comes from the same place: your words sitting on servers you don't control, tied to an account that says who you are. Running a model on your own computer removes the servers. Morse is a web app, so its models run in the cloud, but it's built so the cloud side holds as little as possible. For a model that never touches a network, use a local app; Morse is for the models a laptop can't run.

• E2EE, for the provider row. Your browser checks the secure enclave's hardware proof against versions pinned in the app, then encrypts the message so only that enclave can read it. Morse's servers relay ciphertext they can't read, and the reply is decrypted only in your browser. Open-weight models like GLM, Kimi, and DeepSeek run this way, on enclaves run by third-party providers, not by Morse. It rests on the enclave hardware and attestation working as designed. E2EE chats come with Pro, at $19 a month.
• Anonymous, for GPT, Claude, and Gemini. These are the models no one can download. Our proxy removes your account, email, and IP address and sends the message under Morse's own account, so the provider answers without knowing it's you. It may keep the message for a limited time to check for abuse, under its own policy, so leave identifying details out of the text.
• Private, for everyday chats. Open-weight models plus xAI's Grok, on hosts that agreed to keep nothing: Morse has zero data retention agreements with its inference, search, and other providers. Where a provider confirms zero retention on each reply, Morse checks it and refuses a reply that doesn't.

The device row is covered in every mode. Your chat history is stored only in your browser, encrypted, and locked when you sign out. Unlocking it takes two keys, one that never leaves your device and one our server gives you when you sign in, so a copy of your browser data alone can't be read. That's the part a plain local app leaves to you.

The court row follows from it. Morse's servers don't keep your messages or log what you write, so as our privacy policy says, we usually have very little to give authorities. We don't train on your chats, sell your data, or show ads. Clearing the browser or losing the device deletes the history for good, and we can't recover it, so keep a copy of anything you need. For the employer row, use Morse on your own device, not the work laptop.

Try a chat from the Morse home page without an account, or start free. E2EE chats come with Pro, at $19 a month, with a lock on each reply that shows each check that passed.