Local AI engine: Ollama setup and recommended models

What it does

The app probes the local engine and reports what it measured: whether it answered, its version, the round-trip latency, and the models it declares — each with its weight on disk, parameter count and quantization when the engine publishes them. An engine that is switched off comes back as unreachable with the raw message, which is a result rather than a diagnosis: it does not say whether the key, the network or the service is at fault.

Requests go through the application's Rust backend rather than the web view, for two reasons: the content security policy does not open a local HTTP port from the front end, and networking does not belong there.

Because the list comes from the engine, a model you do not have never appears as selectable. That leaves one question the engine cannot answer — what to install when you have nothing — which is what the suggestion list is for.

Recommended models

How to set it up

• Install an engine yourself — Ollama or LM Studio — and leave it running. Senzoukria never installs the engine; it only talks to one already present.
• Pull a model. The picker shows the exact command to paste, in the form ollama pull qwen2.5-coder:7b, and the Install button asks the engine you already have to fetch it. Progress is reported as the engine's own step names, because Ollama counts bytes per layer and a percentage would fall back to zero on each new layer.
• A suggestion disappears from the list as soon as the model is installed — it then shows up in the real list of installed models instead.
• Set the context window in Settings → AI. The default of 8k fits the assistant's prompt, the market context and a short conversation; Ollama's own default of 4 096 is not enough for that.

Limits and pitfalls

• A local model is a free fallback, not an equal: it is clearly less capable than Claude and will miss things Claude catches. In exchange, nothing you write leaves the machine.
• Local engines receive plain text, so screenshots never reach them whatever the model claims to support.
• Embedding models such as nomic-embed-text are listed by Ollama alongside conversational ones but cannot answer at all; the picker flags them.
• Sizes are the download in the default quantization, and free RAM of the same order is needed: a 5 GB model does not run comfortably under 8 GB of free memory.
• The suggestions favour 7B and above for indicator code, where smaller models tend to produce plausible code that does not run; the single 3B entry is there for machines that cannot hold more.