Skip to main content
← Back to Blog
LLM Tech7 min read

Liquid AI’s LFM3 Moves Tool-Using Agents Toward the Edge

Liquid AI’s LFM3 puts six open-weight multimodal models near two billion parameters on edge devices, changing where agent memory and control can live.

Liquid AI released LFM3 on October 27, 2026, a family of six open-weight models built from a 1.7-billion-parameter backbone. The release covers text, vision, audio, multimodal, and decision models, all roughly two billion parameters or smaller. The company says the text model can occupy about one gigabyte at four-bit precision, while every model can be adapted on a single GPU.

That combination matters more than another benchmark win. LFM3 is designed to put tool-using agents on phones, laptops, vehicles, and private servers instead of routing every instruction through a frontier cloud model. Its weights are available now on Hugging Face, with formats for llama.cpp, MLX, ONNX, and other common runtimes. If the models hold up outside Liquid's evaluations, they make local agent architectures materially easier to build.

The evidence does not support calling LFM3 a general replacement for large models. Liquid's own model card warns against agentic coding and complex multi-step reasoning. Its strongest results are concentrated in selected tool-use and browser tasks, while larger rivals remain ahead on other agent and knowledge evaluations. The practical story is narrower and more useful: small models are becoming credible specialists, and that changes where an agent's state, memory, and permissions can live.

One backbone, six specialized models

The family starts with LFM3-1.7B-Base. Liquid describes a 24-layer tri-hybrid architecture with 13 gated short-convolution layers, five sliding-window attention layers, and six global-attention layers. The model was pretrained on a company-reported 22 trillion tokens, including a final 3.5 trillion-token stage weighted toward agent traces and tool calls. Its advertised context window is 256,000 tokens.

Specialized branches add text instruction following, vision and video input, streaming audio, combined image-video-audio input, and two decision models intended to return calibrated answers. This is not one model dynamically becoming every modality. It is a shared foundation used to produce separate checkpoints, each with its own encoders, decoders, and operating envelope.

The architecture reflects an edge-computing tradeoff. Convolutions and limited-window attention reduce the cost of processing long sequences, while a smaller number of global-attention layers preserve access across the full context. Liquid says it searched across 14 device and backend combinations, from a Raspberry Pi 5 and phones to server GPUs. Those are company claims, not independent performance guarantees, but the public checkpoints make them testable.

The benchmark table argues for routing, not replacement

Liquid reports unusually strong results for a model this small. Its post-trained text model scores 71.39 on BFCL V4 function calling and 38.42 on WebArena-Verified in the company's setup. VentureBeat's review of the published table also shows the limits: LFM3 trails Qwen3.5-9B on MCP-Atlas and narrowly trails it on EnterpriseOps-Gym. Liquid told the publication that larger models still lead on agentic coding and knowledge.

Comparisons across agent benchmarks are fragile. Scores can depend on the harness, tool definitions, retry budget, prompt, reasoning allowance, and whether failures are repaired by surrounding software. Liquid did not provide VentureBeat with confirmation that every comparison used equivalent tools and reasoning budgets. The release also names Mercedes-Benz and Shopify as pre-release evaluation partners but does not publish independent partner results.

Teams should therefore read the model card as a deployment hypothesis. Test the exact workflow, including malformed calls, stale context, permission errors, network loss, and escalation. A small local model may be excellent at classifying a request, extracting fields, or selecting a known tool while remaining unsafe for an open-ended task. The right architecture may route routine steps locally and send ambiguous or high-risk cases to a larger model.

A long context window is not persistent memory

LFM3's 256,000-token window is useful, but it does not make an agent remember across restarts. Context is input supplied for one inference. Persistent memory is state that survives sessions, carries provenance, can be corrected or deleted, and is retrieved according to the current task. Packing more history into a prompt can increase latency and make old instructions compete with relevant evidence.

Edge deployment makes that distinction operational. When inference happens on a user-controlled device, memory can remain nearby as an encrypted store rather than being copied into a provider's conversation history. The model can retrieve a small working set, act, and write a structured result back. That can reduce data exposure and offline dependence, but only if the application defines retention, provenance, conflict handling, and synchronization explicitly.

Local does not automatically mean private. A device agent may still call remote tools, sync state, load untrusted documents, or expose memory through logs. Open weights also do not mean unrestricted software: the checkpoints use Liquid's LFM Open License, so adopters need to review its terms rather than assume an OSI-approved open-source license. The security boundary is the complete system, not the location of the matrix multiplication.

What changes for agent builders

The release strengthens a split that is already visible in production systems. Frontier models remain the best choice for difficult reasoning and broad knowledge. Small specialists can sit closer to sensors, private data, and repetitive tools. A local model can decide what to retrieve, normalize observations, or execute a constrained workflow without sending the full memory record to a cloud endpoint.

That design also makes orchestration more important. Every handoff needs an identity, a versioned model and prompt, a record of retrieved evidence, and a clear account of which component changed state. Otherwise a cheap fleet of edge agents produces fragmented histories that no operator can reconstruct. The smaller the model, the more its reliability depends on good retrieval, narrow permissions, deterministic validation, and escalation paths.

LFM3 is consequential because it packages those possibilities into downloadable, multimodal checkpoints small enough for ordinary hardware. It does not prove that a two-billion-parameter model can safely run an enterprise workflow end to end. It does show that the local part of a hybrid agent system can do more than wake-word detection or simple classification. The next test belongs to builders: whether specialized edge agents can remain accurate, inspectable, and coherent after the benchmark ends.

Sources

Your AI remembers everything. Everywhere.

Unimatrix gives you a shared, durable memory layer across Claude Desktop, Cursor, ChatGPT, and Gemini. Setup in 2 minutes. Free and paid plans available.

Looking for developer resources?

Browse our catalog of 500+ tested AI prompt profiles covering DevOps, data modeling, agent behaviors, and API wrappers. Have a prompt to share? Submit your own for review by our librarian to be featured. Completely free, no registration required. Browse prompt libraries →

Keep reading