xllama
A local LLM chat, image-diffusion and on-device training app for the Xbox Series S and X, built as an architecture exercise under a console's UWP sandbox and 10 GB RAM ceiling.
xllama runs local LLM chat, image diffusion and on-device model training on an Xbox Series S or X, dispatching between a CPU path and DirectML depending on the workload, and exposes an OpenAI and Ollama-compatible LAN endpoint that is off by default. It is framed by its author as an architecture showcase under real constraints: the UWP app model, console hardware, and about 10 GB of usable RAM.
On-device training supports a partial fine-tune lane, a host-side parameter-efficient fine-tuning lane, and serving a merged GGUF checkpoint, alongside chat with key-value cache reuse and thinking models, and a headless benchmark mode for tokens per second, memory and disk bandwidth. It installs from a pre-built package through a CI release or by building from source on Windows, and is written in C++.