Is the default chat model running locally?

[daffodilistic]

Hi I'm not quite sure of the answer given by the default model that is downloaded on install, as it says the execution is not on my computer but on the cloud server? Kinda odd since I had to download the 2.8GB model in the first place? Screenshot attached.

[abitrolly]

Maye it doesn't know, because system prompt doesn't say 'Hey, you are a local model now`.

Need to find out where the system prompt it.

[abitrolly]

The system prompt or, rather, instructions - https://github.com/orgs/janhq/discussions/8933

Is this chat running locally currently?

Looks like a generic test for any model + instructions.

[wellkilo]

The downloaded 2.8 GB model can still be running locally even if it says it is in the cloud. That sentence is model-generated text, not a runtime attestation. A model only sees the messages/tools Jan gives it; it does not inherently know which provider process produced its tokens.

Verify the selected provider rather than asking the model:

1. Open the model picker for this thread. If the provider is llamacpp (or mlx on a supported Mac), Jan is using its local engine. A remote provider such as OpenAI/Azure/Gemini is not local.
2. Disconnect the network after the model is loaded and send a harmless prompt. A successful response is a practical local-only check; web search/MCP tools must be disabled for this test because those may need the network independently.
3. On Linux, inspect the Jan process connections while generating:

ss -tpn | grep -i jan

Jan's llama.cpp extension starts an OpenAI-compatible endpoint on loopback (http://127.0.0.1:<port> / http://localhost:<port>/v1) and sends generation requests there. Current source: local engine props request and local inference base URL.

Why the answer was wrong: the ordinary Assistant instructions define behavior/persona, not a trustworthy statement of deployment topology. Current Jan Agent even uses a generic default identity (You're currently running on Jan agent harness) when no Assistant instructions are supplied; it does not tell the model “you are definitely local.” See the context builder.

So the screenshot is evidence of model hallucination, not evidence that the downloaded model was ignored. The provider shown in Jan's UI and the observed network/process path are the reliable checks.