Skip to main content
NOVEMBER, WEDNESDAY 18, 2026 · SAN FRANCISCO

On-Device
Agents.

The next frontier of AI is closer than you think.

A technical mini-conference exploring the models, silicon, and systems that put useful agents on the devices we already own.

BROUGHT TO YOU BY

6:00–9:30 PM PST

San Francisco, Liquid AI HQ

ONE EVENING. THE COMPLETE STACK.
01

Model

Efficient architectures. Open weights. Models designed for local, agentic workloads.

02

Compute

NPUs, Qualcomm, and Apple silicon. Quantization, memory, and power. Making inference work on-device.

03

Benchmark

Measure the deployed system. Quality, latency, and efficiency under real constraints.

The Speakers

Researchers, engineers, and founders.
In the same room. In conversation.

Model
Compute
Benchmark
Mathias Lechner

Liquid AI

Mathias Lechner

Co-founder & CTO

Efficient models for local intelligence. A conversation about the architectures and practical constraints shaping the next generation of on-device agents.

Henry Ndubuaku

Cactus Compute (YC S25)

Henry Ndubuaku

Founder

8–29 MB foundation models and inference for tiny devices.

Alan Zhu

Qualcomm Nexa

Alan Zhu

Gen AI Product · Previously CPO of Nexa AI

NPUs, hardware acceleration, and the realities of deploying AI on devices. A conversation about running fast, fully on-device inference on the Qualcomm NPU.

Glenn Sonna

Xybrid

Glenn Sonna

Co-founder & CEO

The runtime layer: one SDK that runs LLMs, speech-to-text, and text-to-speech inside apps and games, on the device, with a fallback to the cloud when needed.

Yeon Kim

ZETIC

Yeon Kim

Founder & Technical Lead

Testing, benchmarking, and optimizing models for real on-device deployment through hardware-aware inference.

WEDNESDAY, NOVEMBER 18

Agenda

  1. 6:00PM

    Doors open

    Food, drinks, and conversations with the people building the stack.

  2. 6:30PM

    Intros

    Host introductions and a quick round of intro from attendees.

  3. 7:00PM

    Socratic Dialogues 🔥

    Focused Socratic dialogues around edge AI, on-device agents, and small language models.

  4. 7:40PM

    Technical talks & Q&A

    Five 15-minute technical talks across the stack: models, compute, and benchmarks. Each followed by 5 minutes of Q&A.

    1. Model
      1. Efficient architectures for local intelligence

        Mathias Lechner · Liquid AI

        The model designs and practical constraints shaping the next generation of on-device agents.

      2. Foundation models in 8–29 MB

        Henry Ndubuaku · Cactus Compute (YC S25)

        Tiny foundation models and the inference engine that runs them on small devices.

    2. Compute
      1. Inference on the NPU

        Alan Zhu · Qualcomm Nexa

        Hardware acceleration and the realities of fast, fully on-device inference on the Qualcomm NPU.

      2. One runtime for on-device AI in every app

        Glenn Sonna · Xybrid

        Running LLMs, speech-to-text, and text-to-speech inside apps and games, and deciding when a request stays on the device.

    3. Benchmark
      1. When the Benchmark Changes the Code: Graph and Kernel Engineering for On-Device Agents

        Yeon Kim · ZETIC

        Testing, benchmarking, and optimizing models for real on-device deployment through hardware-aware inference.

  5. 9:20PM

    Closing remarks

    Thanks, takeaways, and what comes next.

  6. 9:30PM

    Keep the conversation going

    Networking at a nearby bar for anyone who wants to join us.

The Guests

80 confirmed guests will join this event from top companies working on edge AI, on-device agents, small language models, robotics, and more.

Where they work

  • Meta
  • Google
  • NVIDIA
  • Apple
  • Amazon
  • TikTok
  • Hugging Face
  • Verizon
  • Walmart
  • Ford
  • Lucid
  • Nutanix
  • UC Berkeley
  • Together AI
  • MacPaw
  • Artificial Analysis
  • Firecrawl
  • Runlayer
  • Prism ML
  • Restart
  • nnopt
  • RunAnywhere

What they work on

  • 35%build agents
  • 30%ship AI on device
  • 20%make models smaller & faster
  • 15%are founders

Built in this room

  • MobileLLM

    Meta · ICML 2024
  • Minitron

    NVIDIA · NeurIPS 2024
  • KVQuant

    UC Berkeley · NeurIPS 2024
  • SmolVLM

    Hugging Face · 2025
  • INT8 Whitepaper

    Google · 2018
  • AIMET

    Qualcomm
  • ONNX Runtime

    Microsoft
  • PyTorch

    Open source

Sponsors

Presenting sponsor

Liquid AI builds efficient foundation models designed to run where the data is: on phones, laptops, vehicles, and edge devices. Our host and presenting sponsor for the evening.

Venue, food, and drinks generously provided by Liquid AI.

Speaker companies

Get your product in front of the 80 people building on-device AI.

Sponsors meet a curated room of engineers, researchers, and founders who choose the models, runtimes, and hardware their products ship on. Leave with conversations with the teams you want as users, partners, and hires.

Become a sponsor

Invite only

Attendance is by invitation. Ask to join on Luma — every request is reviewed to keep the room technical.

GET IN THE ROOM

Build what
comes next.

For AI researchers and engineers working on small local models and on-device agents. Participation is curated to keep the conversation technical and high signal.

Hosted by Federico Ulfo, Viviana Marquez, …

Part of Frontier Stack — bringing frontier labs and the AI ecosystem into conversation with builders.

Explore the initiative