LocalAI has announced several new AI models, including the 'q2.5-ms-mistoria-72b-v2', a 72 billion parameter model designed for storytelling and intelligence. Users can install it using the command 'local-ai run q2.5-ms-mistoria-72b-v2'. Additionally, LocalAI introduced 'llama-3.1-tulu-3-8b-SFT', an instruction-following model, and 'eva-qwen2.5-72b-v0.2', which specializes in role-playing and storywriting. The 'llama-3.1-tulu-3-70b-dpo' model was also launched, showcasing state-of-the-art performance across various tasks. Furthermore, a lightweight model, 'Llama-SmolTalk-3.2-1b-instruct', was introduced for efficient text generation and conversational AI. HuggingFace has expanded its Inference Endpoints to support deploying llama.cpp-powered instances on CPU servers, which could lead to more affordable cloud LLM options as new AI-friendly instruction sets emerge. This development is seen as a significant step toward making large language models more accessible.