Model Orchestration
Model Orchestration automatically selects models based on their capabilities, so you do not need to decide which model best fits each task or workload.
The main orchestration model, sl-mini, routes tasks based on their required capabilities.
Select one or more models in the model selection modal (Ctrl+Shift+M). When you select multiple models, Auto Mode considers the available resources, task, and workload to choose the best model for each step.
How it works
For example, suppose you want to extract information from an ID card and have selected the following models:
When you attach the ID image and ask to extract key information, the orchestrator follows this flow:
Example orchestration timeline
| Time | Stage | What happens |
|---|---|---|
0.00s | Prompt received | SmartLoop receives your request. |
0.04s | Attachments resolved | The attached ID image is prepared for processing. |
0.04s - 7.41s | Image processing | qwen2-vl-2b reads the image and extracts visual information. |
7.41s - 9.79s | Task planning | sl-mini plans the response, selects tools, and loads the vision model. |
9.79s - 12.73s | Vision analysis | Gemma 4 E2B It (4.6B) analyzes the image for the requested information. |
12.73s - 14.93s | Response generation | The selected model prepares and generates the final answer. |
14.93s | Answer delivered | The response completes with the key information extracted |
The image is first processed by the specialized vision model, then the orchestrator selects the best available model to analyze the result and produce the response.
Open the orchestration log (Ctrl+Shift+L) to inspect each step and understand how the request was processed.
Memory Management
To select the most suitable model, the orchestrator evaluates the memory available on your device alongside the requirements of the task. On Apple Silicon Macs, this means checking available unified memory; on NVIDIA GPUs, it checks available VRAM.
SmartLoop uses Vulkan to support a wider range of hardware. When a dedicated GPU is unavailable, it can use the integrated GPU memory provided by many modern Intel and AMD processors. The available memory determines which models can run reliably, so selecting fewer or smaller models can help on devices with limited resources.
Auto Mode
Auto Mode chooses the right models for every task, so you can focus on the work instead of model settings. SmartLoop matches each step to the capabilities it needs, then validates the workflow as it runs to deliver the strongest possible result.
Next Steps
We generally use q4_k_m quantization to run models efficiently with minimal quality loss on resource-constrained devices. We will adopt improved compression methods as they become available and stable.
We will add new models based on use cases and capabilities. To request a model, contact us at support@smartloop.ai.