First-Time Setup¶

On the first launch, Backend.AI GO opens a six-step onboarding wizard.

1. Language
2. Accelerator choice
3. Models directory
4. Optional first model download
5. Optional cloud provider setup
6. Completion

Supported interface languages¶

The current wizard supports these UI languages:

• English (en)
• Korean (ko)
• Japanese (ja)
• Spanish (es)
• French (fr)
• Traditional Chinese (zh-TW)
• Simplified Chinese (zh-CN)

Step 1: Language¶

Choose the interface language first. The rest of the wizard updates immediately to match your selection.

Step 2: Accelerator choice¶

Backend.AI GO detects the host and recommends an accelerator family.

• On Apple Silicon, this usually points you toward Metal-capable local engines and MLX support.
• On NVIDIA, AMD, and Intel systems, the recommendation reflects the available local acceleration path.
• You can keep the recommended option or switch to a manual choice.

On Apple Silicon, once the hardware is detected, the step also offers an Inference engine:

• llama.cpp (the default) runs GGUF models and supports most models and features.
• MLX is faster on Apple Silicon but has limits: it runs MLX-format models only, some features such as embedded MTP speculative decoding and the llama.cpp hardware settings are not available, and some depend on the engine version. Choosing it installs the mlxcel engine after setup. See MLX.

The engine is fixed once the first model starts downloading, since that download is for the chosen engine.

The choice is recorded in settings and acted on when the wizard finishes. If no installed engine already serves the chosen family and the app carries no built-in engine, the matching llama.cpp engine is downloaded in the background after setup. A headless server install is the case where this matters: it ships no embedded engine, so without the download it would finish onboarding with nothing able to run a local model.

Onboarding never waits for that download. Progress appears in the notification center and in the header download indicator, and a failure is reported there with a link to Engines instead of blocking the wizard. The step 6 summary states which of the outcomes applies: an engine will be installed, one is already installed, the built-in engine will be used, or nothing will be installed automatically.

Step 3: Models directory¶

Choose where downloaded local models should live. Leaving the field empty tells the server to use its default app-data path.

Default app-data roots are:

The older ~/.backend-ai-go/ path is legacy-only. Current desktop installs use the ai.backend.go app-data directory.

Step 4: Optional first model download¶

You can finish onboarding without downloading a model, or let the wizard queue a recommended first model.

With MLX chosen in step 2, each pick is the MLX version of the same model (a repository of several files, slightly larger than the GGUF file). When the download finishes during onboarding, the step dots at the top of the wizard briefly give way to a "finished downloading" line instead of a notification that would leave the wizard.

The wizard can also offer a companion extraction model. This second download is optional and is intended for workflows that use extraction or utility-model paths.

After onboarding, you can always download more models from Models > Browse.

Step 5: Optional cloud provider setup¶

You may optionally register API keys for supported cloud providers during onboarding.

The wizard currently supports:

• OpenAI
• Anthropic
• Gemini

This step is skippable. You can add or edit providers later from Settings > Providers.

Step 6: Completion¶

The final screen summarizes the selections and sends you to Home, not directly to the power-user dashboard.

From there you can:

• open Models > Browse to download more models
• open Models > Local models to run downloaded models
• start a conversation from Home, and attach a working folder in the composer to run it as Cowork
• open Engines to install additional local or container engines