Weights are learned numerical parameters
When a developer releases model weights, other people can download those parameters and run the model in compatible software. “Open weights” does not automatically mean the training data, training code, or every component is available under an open-source license. Read the specific model’s license and documentation.
Local is a place, not a privacy guarantee
Local inference means the model is running on your device. That can keep the inference itself off a hosted model service. But browsing tools, plugins, telemetry, cloud synchronization, and app settings can still create network activity. Inspect the entire workflow before using private material.
You also need enough memory and storage. A compressed, or quantized, model can reduce memory use, with tradeoffs that depend on the model and method. Parameter count, context length, and the runtime all affect the experience.
How our recorded examples were made
Our recorded comparison examples used a Qwen model already loaded in LM Studio on a Mac. The saved run identifier is qwen/qwen3.8-27b, with a loaded context of 16,384 tokens. These are observations about one setup, not a hardware recommendation. The public comparison pages replay saved outputs; opening them does not download or run that model on your device.
Treat a model download like software
Check who published the files and who converted them into your chosen format. Use the publisher’s documentation, read the license, and avoid running installation commands you do not understand. A model card, a popular download count, or a matching file hash is useful evidence about provenance or integrity, but none proves a download is harmless.
Some older checkpoint formats use Python pickle, which can execute code when loaded. Safetensors avoids that particular behavior; the runtime, libraries, and any custom model code still need your trust. A safer file format is one part of a safer setup.
Start with what you have
Before downloading anything large, read the model card, check the app’s supported formats, and compare the memory requirement with available memory. Try a short, non-sensitive task first. Local models can be a useful hobby and a practical tool, but setup time and electricity are costs too.