Bringing AI Agents to Your Local Hardware 🕵️‍♂️

For the past few years, the narrative around generative artificial intelligence has been dominated by massive cloud-hosted models. While huge models residing in distant data centres deliver impressive results, they also introduce latency, soaring operational expenses, and privacy concerns for enterprise workloads. Liquid AI is shifting that paradigm with the release of its LFM3 model family. Designed specifically for specialised intelligence and on-device execution, LFM3 proves that you do not need tens of billions of parameters to run sophisticated browser agents, interpret voice commands, or perform complex visual reasoning.

Redefining the Architecture for Edge Efficiency

At the heart of this release is a shared 1.7-billion-parameter backbone, expanding into 2B-class models tailored for distinct modalities. Rather than relying purely on standard transformer blocks, Liquid AI engineered a novel tri-hybrid architecture. By interleaving gated short convolutions, sliding-window attention, and global grouped-query attention, the architecture achieves high expressivity while drastically cutting memory overhead.

According to Liquid AI’s official announcement, this structural redesign allows the LFM3-1.7B text model to outperform older, larger models while using 35% fewer parameters. Furthermore, by expanding its hardware testing across 14 different mobile and desktop platforms—ranging from Apple M-series chips and AMD Ryzen AI processors to Qualcomm Snapdragon hardware and smartphones—Liquid AI has built a model family capable of running at lightning-fast inference speeds right on company-controlled hardware.

Smart Browser Automation and Agentic Workflows

One of the most compelling aspects of the flagship LFM3-1.7B text model is its post-training for agentic harnesses like OpenClaw, Hermes Agent, and Pi. Out of the box, it supports a massive 256K-token context window on downloadable builds, allowing it to parse extensive tool traces, execute multi-step web research, triage inboxes, and manage schedules seamlessly.

As explored in VentureBeat’s enterprise analysis, LFM3-1.7B punches well above its weight class on real-world automation benchmarks. On tasks like WebArena-Verified and AutomationBench, it comfortably outperforms larger models such as Qwen3.5-9B. Because the model footprint sits around 1 GB at 4-bit precision, organisations can run dozens of local background agents simultaneously without paying per-token cloud charges or exposing sensitive operational workflows to third parties. For developers looking to inspect model parameters, benchmark figures, and prompt templates, full specifications are available in the LFM3-1.7B documentation.

Multimodal Power: Vision, Speech, and Decision Models

Local intelligence is not restricted to plain text. The LFM3 family spans six distinct models branching from the core backbone, covering vision, speech, omni-modal inputs, and structured decisions:

• LFM3-VL-2B (Vision): Adds native video understanding, spatial reasoning, and GUI interpretation. It handles document OCR, charts, and visual troubleshooting locally with remarkable accuracy.
• LFM3-Audio-2B (Speech): Features an advanced audio encoder that begins processing speech before the user even finishes speaking, enabling voice interaction response times as fast as 50 milliseconds.
• LFM3-Omni-2B (Omni): Combines visual, speech, and text inputs into a unified hands-free assistant that can listen to spoken instructions while inspecting images or video feeds.
• d1-2B and d1-Omni-2B (Decision): Purpose-built models that evaluate inputs and return structured JSON outputs, numerical estimates, and confidence scores for task routing and invoice classification.

Whether it is a field technician inspecting damaged equipment via a smartphone camera or a warehouse worker using voice commands to update inventory hands-free, these compact models bring real-time AI to where work actually happens.

Enterprise Deployment and Customisation

For enterprise teams, the economic argument for local deployment is clear: predictable costs, zero data retention, and independence from cloud availability. Crucially, every model in the LFM3 family can be fine-tuned on a single GPU, allowing businesses to adapt models to domain-specific datasets in just a few hours.

For organisations that prefer cloud infrastructure alongside on-device deployment, hosted endpoints are available through the Liquid AI Console. The API pricing is remarkably competitive—starting at $0.08 per million input tokens and $0.20 per million output tokens for the text model—making it an attractive choice for hybrid deployments.

A Pragmatic Step Toward On-Device Intelligence

Liquid AI's LFM3 release demonstrates that smaller, highly specialised models can deliver genuine operational value. By moving browser automation, speech processing, and visual analysis off remote servers and onto local laptops, phones, and private infrastructure, LFM3 opens up practical, secure AI deployment for modern enterprises.

Disclaimer: This article is provided for informational purposes only, mistakes may be made, and it's not offered or intended to be used as legal, tax, investment, financial, or any other advice.