Mima uses local small language models to manage social connections

Rachel Lee Neighbors, formerly of the React team and W3C, is building Mima, a local-first client for managing social connections across X, LinkedIn, Blue Sky, and Mastodon. The application runs small language models locally to avoid sending private data to cloud servers, citing privacy and energy costs. Neighbors initially used Claude Sonnet for summarizing conversation threads but switched to Llama 3.2b after setting up evaluations with the Phoenix library. She found that Llama 3.2b performed as well as Claude Sonnet for local summarization with lower latency than Gemma 4. The system uses multiple specialized small models for tasks like toxicity detection and moderation, all of which are opt-in downloads rather than default installations. Neighbors notes that local models can sometimes be slower than cloud round-trips depending on hardware, such as Mac chips, but emphasizes that keeping data on-device prevents external access to personal information. The project is launching with a pop-up soft launch in San Francisco. Neighbors argues that local-first, small-model architectures are superior for consumer-facing AI agents because they protect user privacy and reduce the energy and financial costs associated with large-scale cloud inference.