Project A.E.S.I.R. is an experimental native-Mojo engine for running local AI models directly on your own hardware, and it is evolving from a working low-level inference engine into a practical backend that ordinary AI applications can eventually use.
Right now, A.E.S.I.R. can already run real GGUF language models through native Mojo on both CPU and NVIDIA GPUs, hold conversations with them locally, manage and inspect models, download models from Hugging Face, and serve them through working Ollama- and OpenAI-compatible APIs.
Development is moving quickly, with plans to expand A.E.S.I.R. across many more model families, quantization formats, GPUs, CPUs, NPUs and edge devices, while adding richer application compatibility, smarter heterogeneous hardware use, advanced memory and inference optimizations, multimodal AI, and eventually a foundation for more ambitious local machine-cognition systems.
Developers and testers are very welcome to join the forge, whether by adding support, improving the runtime, testing new model families and quantizations, or helping get additional hardware backends properly tested, documented, and proven with reproducible evidence.