Models

Source: src/renderer/src/components/ModelsInterface.tsx

Overview

Browse, download, and manage LLM models. Models are stored locally in ~/.silicon-studio/models/.

Model Sources

The Models page lists models from multiple sources:

1. Embedded / Recommended — Specific native models shipped via configuration (es. fabriziosalmi/nanocoder). NanoCoder is an embedded tool-calling agent fine-tuned so that it does not rely on <think> xml tags when reasoning, outputting everything purely as plain inline-text. Automatically pulled on system launch if missing!
2. Hugging Face — Browse and download MLX-compatible models. Filtered to mlx-community and similar repos.
3. LM Studio — Auto-discovers models in ~/.lmstudio/models/.
4. Ollama — Auto-discovers models in ~/.ollama/models/.
5. HuggingFace cache — Auto-discovers models in ~/.cache/huggingface/hub/.
6. Custom — Register any local model by path.

Model Registry

All known models are tracked in ~/.silicon-studio/models.json. Each entry contains:

Field	Description
id	Unique identifier (Hugging Face repo ID or generated UUID)
name	Display name
size	Parameter count (e.g., "1.7B", "7B")
architecture	Model architecture (qwen2, llama, mistral, etc.)
downloaded	Whether files exist locally
local_path	Absolute path to model directory
is_finetuned	Whether this is a fine-tuned adapter
url	Hugging Face URL (if applicable)

Operations

Download

Click the download button on any Hugging Face model. Downloads run in the background. Progress is not currently tracked in the UI — the model appears as "downloaded" when complete.

Delete

Removes the model files from disk and marks it as not downloaded in the registry.

Register Custom Model

Provide a name, local path, and optional Hugging Face URL. The model is added to the registry immediately.

Scan Directory

Point to a directory containing model folders. All valid MLX models found are registered automatically.

Loading and Unloading

Only one model can be loaded in memory at a time. Load a model by:

• Clicking "Load" on the Models page
• Using the model dropdown in the top bar

The top bar shows the currently loaded model name with an Eject button. Loading a new model automatically unloads the previous one.

Backend implementation: backend/app/engine/service.py — calls mlx_lm.load() to load the model and tokenizer into MLX memory. KV-cache quantization (4-bit or 8-bit) can optionally be enabled at load time; it is applied during generation via the kv_bits= parameter to stream_generate, not during the load() call itself.

Model Card

Clicking a model opens a split-view detail panel showing:

• Model metadata (size, architecture, path)
• Actions (download, delete, load)
• Status indicators