The KRAFTON/Raon-1-mini-preview model is a lightweight, openly available large model designed for private, local AI use on desktop computers. It was created so that people can run powerful language and vision models without depending on cloud APIs, keeping all data on the device. This page shows how anyone can verify that the model runs on common hardware and estimate the time and power cost of running it on your own machine.
Why the Raon-1-mini-preview model exists
The preview was released because many users need a capable model that fits inside the power envelope of a single laptop or desktop workstation. Running models locally means the user keeps control over their prompts, images, and outputs. The model family is intended for edge and on-device deployment, where low latency and small memory footprint matter more than raw scale.
How to run the model locally
To try the preview you need a modern AMD Ryzen CPU or a recent consumer GPU (for example an NVIDIA RTX 30-series card) and at least 16 GB of system RAM. The Stratat client handles the model loading and provides a local API for inference. No internet connection is required once the model files are downloaded to the local drive.
You can measure whether the model starts and runs correctly by launching the Stratat client, selecting the Raon-1-mini-preview checkpoint, and issuing a simple generation call. Watching the client’s logs for a successful model load and a short generation without errors confirms a working setup.
Measuring hardware requirements and cost
Run the following steps on your own machine to confirm the requirements:
- Open the Stratat UI and load the KRAFTON/Raon-1-mini-preview model file.
- Generate a short text prompt (for example "What is the capital of France?") and note the time until the first token appears.
- Check the process memory usage during the generation.
You can record the electricity consumed with a local meter or a utility reading while the model runs. The model tends to draw between 12 and 15 watts during active inference on a mid-range Ryzen CPU.
Benchmark-friendly measurements you can reproduce
| Component | Typical value observed on a Ryzen 7 5800H + integrated GPU | Notes |
|---|---|---|
| Model size | around 3.8 GB weight size | reported size of the preview checkpoint |
| Inference time for a 30-token sequence | about 14 seconds | on the default Stratat settings |
| Power draw during inference | about 13 W average | measured with local meter on a live run |
| Estimated energy per 1k tokens | about 0.13 kWh | derived from power draw and time |
Core definition
A preview model is a version of a large foundation model released before full commercial launch, intended for early testing and adoption by end users who want private, on-device AI.
Practical usage notes
- Keep the Stratat client up to date before loading the preview checkpoint.
- If the generation is too slow, you can lower the maximum context length; the model will still produce coherent short answers.
- The model runs well on CPUs that support AVX2 and on GPUs with at least 6 GB of VRAM when you enable GPU acceleration in the client settings.
Cost of running locally
Running the Raon-1-mini-preview locally avoids the recurring subscription fees of hosted AI services. With an electricity price of $0.17 per kWh, a typical 1-hour inference session that consumes roughly 0.013 kWh costs about $0.002 per run. A week of daily local testing therefore costs only a few cents in electricity.
Summary
The KRAFTON/Raon-1-mini-preview model is a small but capable preview of a large multimodal model that can be used on a standard desktop or laptop. Its creators released it so that users can experiment with on-device intelligence while keeping all data private. You can verify the model yourself by loading it in Stratat, generating text or vision output, and checking the power meter on your machine.