The developer introduced Mini-AGI with training on 8 GB VRAM

The developer presented Mini-AGI — an experimental model with a continuous learning architecture capable of functioning on consumer equipment with 8 GB VRAM. The project is focused on working with byte-level data and optimizing resources with limited video memory. The key feature of the system is the mechanism of swapping weights from disk, which allows bypassing hardware limitations when performing calculations. The model uses adaptive depth and a system of expert routing (MoE), providing dynamic distribution of computational load during training. At the current stage, the project is positioned as a research experiment, rather than a competitor to frontier models. The source code is available in an open repository, where discussions of architectural solutions and memory optimization methods for local launch of neural networks are also held. The technical implementation of Mini-AGI demonstrates the possibility of implementing complex learning algorithms on affordable hardware. Details of the implementation and current performance metrics of the model are discussed in the professional community on Hacker News. The development of such solutions allows lowering the entry threshold for training neural networks that previously required server power. The use of weight swapping and adaptive layers opens the way to creating more compact and efficient systems capable of retraining in real-time without significant infrastructure costs.