Flutter Nantes #23 - September 24, 2026

# Flutter Nantes Meetup #23 : Edge AI in Flutter: How Far Have We Come?

📅 September 24, 2026 / 19h–22h
📍 Nantes, hosted by SNCF Connect

We're thrilled to welcome Sasha Denisov for our 23rd meetup, in a special talk + hands-on workshop format: "Edge AI in Flutter: How Far Have We Come?". First a retrospective tracing the entire evolution of on-device AI in Flutter — then you'll build your own offline AI agent, live, on your own machine.

Two years ago, when flutter_gemma first shipped, on-device AI in Flutter meant a MobileNet-style vision classifier and a small text-only LLM answering one prompt at a time — no images, no tools, and no memory. Fast forward to 2026, and the landscape is almost unrecognisable: phones now run multimodal open models — Gemma, Qwen, Llama and more — that see images, call tools, retain context, and retrieve from a local knowledge base. Fully offline, with nothing leaving the device.

We'll start with a retrospective across the models, the runtimes powering them, and the Dart packages that brought it all natively into Flutter. LiteRT-LM is the centre of that story — open-sourced by Google in June 2025, it's the same runtime that powers Gemini Nano in Chrome, Chromebook Plus and Pixel Watch, reaching Android, iOS, Web, Desktop and IoT from a single C API, with GPU and NPU acceleration and its own .litertlm format.

flutter_gemma talks to it through dart:ffi: one client, six platforms. We'll put it next to the alternatives — MediaPipe, llama.cpp, ONNX Runtime, and even OS-native models — and be specific about when you'd reach for each.

Using flutter_gemma as a throughline, we'll see how it grew from a thin MediaPipe wrapper around a 2B model into a full modular ecosystem, complete with pluggable engines, embeddings, on-device RAG, and agents capable of calling tools, running skills, and speaking MCP.

Then comes the workshop part: you build it.
Starting from an empty Flutter app, one capability at a time:

• It talks. Load an open model on LiteRT-LM, stream the first tokens on a real device.
• It sees. The same session, now multimodal — an image in, an answer out.
• It acts. Function calling: your Dart functions, called by the model, results fed back.
• It learns new tricks. Skills as plain SKILL.md files — the format Google's own AI Edge Gallery uses — dropped into a running app. No rebuild, no release.
• It remembers. On-device embeddings and vector search over your own data.
• It flies. Airplane mode on, everything still works — because nothing was ever leaving.

Along the way: real apps shipping on this stack today.

What you'll walk away with:

• How the model landscape shifted from single-purpose, text-only tools to multimodal, tool-using models running fully offline
• A working offline agent you built yourself, running on your own device
• A complete map of the Flutter edge-AI ecosystem: LiteRT-LM and the runtimes around it, the flagship packages (flutter_litert, flutter_gemma), and the surrounding plugins for RAG, agents, embeddings, speech and ML Kit
• What's only become possible recently: on-device RAG, multimodal input, and real agent loops calling MCP tools locally
• Where it's all heading, and what's genuinely production-ready today

💻 This is a hands-on session — bring your laptop with a Flutter dev environment set up (and ideally a physical device) to build along. You can also just watch and follow the demo, of course.