Your first local conversation

1. Confirm the workstation

Open the native application and let initialization finish. The top-right state should report the actual runtime condition. In System, review your CPU, RAM, GPU and capture time. A runtime marked ready means the private service is available; no model has necessarily been loaded.

2. Bring an existing model

Open Model Library. Use Add folder for a directory of GGUF files, or Import GGUF for an individual file. Both use native file pickers. Import selects a file to inspect and run in place; it does not copy or convert its weights.

Choose an artifact and inspect its architecture, quantization, file size, context ceiling and warnings. A vision projector is a companion file and cannot be loaded as a standalone chat model. A single shard from a split model is not a complete loadable set in the current core.

3. Make an explicit load plan

Open Load Lab. Select the model, backend, exact context and GPU-layer count. The qualified RegesCore 35B smoke test used 8,192 context / 12 layers / CUDA on the documented test machine. Those values are a reproducible example, not a recommendation for every artifact or PC.

If the selected backend is missing, choose Install backend and wait for completion. Read the memory estimate and warnings, then choose Load model. Loading can take time. Use System logs for actual observations; there is no fabricated percentage or guaranteed load-cancellation control.

4. Generate and stop

When the header reports Model loaded, open Chat and send a short prompt. Text and any engine-provided reasoning stream separately. Stop generation aborts the active request. After it settles, send a fresh prompt to check recovery.

Use a task you can verify independently. A plausible answer is not proof of factual correctness, and a successful arithmetic example is not a broad quality benchmark.

5. Save and release resources

Completed conversations are stored locally and appear in recent chats. Start New chat for a separate conversation. To release model allocations, return to Load Lab and choose Unload current model. Closing RegesARC stops its owned private runtime; it does not stop another program’s server.

If a step fails, preserve the exact error and read troubleshooting. Changing several settings at once makes the cause harder to identify.