Two deployment paths
Backyard AI's web experience sends requests to hosted models, while its desktop application downloads model files and performs local generation. The desktop path can work offline and stores its chat data on the device according to the product documentation. Hosted chat is easier to access across devices but necessarily involves a remote service. Readers evaluating sensitive roleplay should treat these as different data paths, not interchangeable skins over the same runtime.
Characters, context, and lorebooks
Characters contain a display name, real name, character and user personas, scenario material, first message, examples and model instructions. A lorebook inserts an entry only when its keywords occur in recent messages, conserving context compared with permanently including every setting fact. The remaining conversation still competes for the model's finite context window, so older turns can fall out even when the character definition remains.
Local models and hardware
The desktop model manager uses GGUF files and is documented as powered by llama.cpp. It recommends quantized files such as Q4_K_M or Q5_K_M for common consumer systems and automatically configures GPU acceleration across supported hardware. A quantized model reduces memory use but can trade away some output quality; larger parameter counts and context settings also increase memory demand.
Access and limitations
The hosted free allowance and paid tiers differ in available models and maximum context, while the local application is free. Group chat, voice calls, sampler controls and lorebooks are documented features, but model capability depends on the selected backend. The absence of a text filter does not guarantee coherent adult roleplay, and community publishing remains subject to platform rules even when a private local model produces unrestricted text.
Technical context
Backyard AI's documented local path uses GGUF and llama.cpp; the dossier separates frontend, weights, template, cache, and execution layers.