Nanbeige/Nanbeige4.3-33B is a 33B‑parameter open model released for on‑device use. Stratat provides a desktop client that streams model weights to the local GPU, letting users run large language models entirely offline. This page shows what the model is for, the hardware it needs, and the steps to verify its behavior yourself.

What Nanbeige4.3-33B is for

Nanbeige4.3-33B is a 33‑billion‑parameter language model released under a permissive license. Its purpose is to give developers a practical foundation model they can run on typical laptop or desktop hardware without sending data to external services. Because the model is open and Stratat is designed for local inference, users keep data private and avoid reliance on remote compute.

Hardware requirements for local runs

The model fits comfortably on GPUs with at least 24 GB of VRAM when running in 16‑bit precision, or on multi‑GPU setups where individual GPU memory can accommodate roughly 8 GB per 8 B of model scale. A modern high‑end consumer card such as an NVIDIA RTX 3090 or an equivalent AMD Radeon (with sufficient VRAM) is recommended for smooth use. The Stratat client will stream the weights to the local device, and the app will report available GPU memory at launch.

How to check the model on your own machine

Run the following steps on your machine: 1. Install the Stratat desktop client from the provided installer. 2. Load the Nanbeige4.3-33B model through the client’s model loader. 3. After loading, open the system monitor to confirm the GPU memory reported matches the model size expectations (e.g., 24 GB for a 33B‑parameter checkpoint). 4. Run a short generation test with the client's command line to ensure tokens are generated without requiring network access.

These steps let you confirm that the model loads and runs locally on your specific hardware.

MetricValueNotes
Parameter count33 BReported model size in published notes
Minimum VRAM for 16‑bit inference~24 GBRough estimate based on model size
Typical consumer GPU that meets requiremente.g., NVIDIA RTX 3090Meets the VRAM threshold

This page is intended for readers who want to evaluate the model themselves.