← All speakers

Bio, Work & Ideas

Adrien Grondin

Conference affiliation: Locally AI · 2026

Adrien Grondin created Locally AI, the Apple-device chatbot acquired by LM Studio in April 2026, and now leads the company’s development of native AI experiences across devices. His work brings language models to iPhones, iPads, and Macs through direct on-device inference or encrypted connections to computers running larger models.

Grondin studied at the Paris engineering school ISEP and developed early iOS augmented-reality and macOS projects before joining Paris startups. He subsequently spent four years developing iOS applications at Match Group, including Meetic and Tinder, while building Locally during evenings and weekends. The project transformed his experience shipping consumer applications into a practical approach to private, accessible AI on Apple hardware.

LM Studio acquired Locally and hired Grondin in April 2026. By June, Locally had become LM Studio’s official mobile application, adding LM Link for encrypted access to larger models running on a user’s computer. Unlike inference performed directly on the phone, these remote sessions require internet access.

  • On-device inference as product design. Grondin curates models that run well on specific Apple devices and weighs download size, memory, response quality, streaming behavior, and differences between phone generations. He demonstrated Gemma 4 generating approximately 40 tokens per second on an iPhone; separately, his iPhone hardware comparison measured Apple’s foundation model at roughly 55 tokens per second on an iPhone 16 Pro and 85 on an iPhone 17 Pro.
  • Native integration through MLX Swift. His applications use Apple’s MLX Swift language-model libraries and compatible weights from the MLX Community. He generally recommends four- to eight-bit quantization, cautioning that more aggressive compression can compromise output quality. The stack supports tool calling, while structured generation remains less mature.
  • Private AI across personal devices. Smaller models can power Apple Shortcuts and lightweight text-processing tasks; larger models can run on personally controlled computers through LM Link. Grondin also demonstrated Liquid AI’s LFM2 8B A1B inside Locally, crediting the upstream MLX port that enabled its Swift implementation.

Read the topics behind these talks

1 conference talk

References