Automated Foundation Model Design (AFMD)

Service · Managed service by Liquid AI

The right architecture, co‑designed on silicon.

Automated Foundation Model Design (AFMD) is a hardware-in-the-loop service that creates bespoke AI architectures tailored for specific devices. It is designed for enterprises needing to maximize performance on edge hardware, delivering up to 5x improvements in speed and memory efficiency.

Languages
English

Category: AI Infrastructure

Deployment
OEM, on-premise, and on-device
Performance Improvement
Up to 5x better RAM/speed
Context Window Support
Up to 128K context (as seen in LFM3)
Hardware Targets
Mobile, Edge, Laptop, Cloud, and NPU silicon
Deployment Models
OEM and On-premise
Privacy Level
Zero data leakage (local training)
  • Up to 5x better RAM and/or speed on average
  • Hardware-in-the-loop architecture search
  • Data never leaves the customer's environment
  • Support for OEM and on-premise deployment

Overview

Automated Foundation Model Design (AFMD) is Liquid AI's proprietary service for creating bespoke AI architectures co-designed with specific hardware. The service is intended for enterprises, such as Mercedes-Benz and Shopify, that need to deploy high-performance intelligence on devices with strict memory, battery, and latency constraints. AFMD works by evolving candidate architectures and scoring them based on task quality and real-world performance measured on the target silicon, including CPUs, GPUs, and NPUs. This hardware-in-the-loop approach allows Liquid AI to deliver tailored models that achieve up to 5x better RAM and speed compared to general-purpose models. The process ensures data privacy by training the selected architecture entirely within the customer's own environment. Liquid AI supports the full lifecycle from hardware profiling to OEM and on-premise deployment.

Architecture Design and Search

  • Hardware-in-the-Loop Search: The AFMD process designs model architectures with the target hardware in the loop, scoring each candidate on downstream task quality, latency, and peak memory.
  • Architecture Evolution: Liquid AI evolves candidate architectures based on the specific memory and latency budgets allowed by the target device and runtime.
  • Component Optimization: The search process selects optimal components, such as sliding-window attention for LFM3, to balance quality and speed for specific workloads.

Enterprise Implementation

  • Private Training: Liquid AI trains the selected architecture for target tasks entirely within the customer's environment to ensure data privacy.
  • Enterprise Deployment Support: The company provides support for OEM and on-premise deployment of the specialized models.
  • Performance Gains: Tailored models deliver up to 5x better RAM and speed on average compared to adapting a general open model.

Published results

  • Up to 5x better RAM and/or speed on average versus general open models

Use cases

  • Consumer electronics manufacturers deploy specialized models on laptops and phones that are optimized for specific chipsets like AMD Ryzen AI or Apple M4.
  • Industrial robotics companies use AFMD to create models that fit the strict memory and latency budgets of embedded edge controllers.

Specifications

Deployment
OEM, on-premise, and on-device
Performance Improvement
Up to 5x better RAM/speed
Context Window Support
Up to 128K context (as seen in LFM3)
Hardware Targets
Mobile, Edge, Laptop, Cloud, and NPU silicon
Deployment Models
OEM and On-premise
Privacy Level
Zero data leakage (local training)

How it works

  1. Profile the target: Liquid AI starts by profiling the target device and runtime to determine the memory and latency budget they allow.
  2. Search on the hardware: The AFMD process evolves candidate architectures and scores them on downstream task quality, latency, and peak memory measured directly on the device.
  3. Specialize for the use case: The selected architecture is trained for the target tasks, ensuring that data never leaves the customer's environment.
  4. Ship and deploy

FAQ

How does the AFMD process work?

The process starts by profiling the target device and runtime to establish a memory and latency budget. Liquid AI then evolves candidate architectures and scores them on task quality and device-measured performance before training the final model.

What are the performance benefits of tailored models?

Tailored models created through AFMD deliver up to 5x better RAM usage and speed, on average, compared to adapting a general-purpose open model.

Is my data kept private during the design process?

Liquid AI ensures that data never leaves the customer’s environment during the specialization and training of the model architecture.

What hardware is supported by AFMD?

Liquid AI supports a variety of hardware including mobile devices (Samsung Galaxy, OnePlus), edge devices (Raspberry Pi), laptops (Apple M4, AMD Ryzen AI), and specialized processors like the Qualcomm Hexagon NPU.

Added by AntSpace from public sources (liquid.ai).

More services by Liquid AI