How Liquid AI’s LFM Models Bypass Transformers to Power Local Agents on Edge Hardware

MIT spinoff Liquid AI has released its new LFM 2.5-2.6B models on Hugging Face, marking a major milestone in the race to deploy fast, private, local agents directly on consumer hardware. By abandoning the dominant transformer architecture in favor of its proprietary Liquid Foundation Models (LFMs), the startup is challenging the industry assumption that edge AI requires heavy-duty quantization or massive hardware compromises.

The Architecture Shift: Beyond the Transformer Bottleneck

For the past half-decade, generative AI has been synonymous with the transformer architecture. But transformers suffer from a fundamental flaw at the edge: the quadratic complexity of self-attention. As the context window grows, memory consumption spikes, making local agent workflows—which require keeping deep history and chain-of-thought steps in active memory—prohibitively expensive for edge devices.

This is where Liquid AI, founded by AI pioneers Ramin Hasani and Daniela Rus, steps in. Their Liquid AI LFM architecture relies on continuous-time state-space representations and adaptive dynamical systems. Unlike transformers, which process tokens in discrete, rigid steps, LFMs process sequential data as continuous streams. This architectural shift yields linear computational complexity and a dramatically smaller memory footprint, allowing the models to process vast context windows with a fraction of the RAM required by traditional architectures.

How LFM 2.5-2.6B Compares to Llama and Gemma

To understand the disruption, one must compare the new LFM 2.5-2.6B models to the reigning champions of the edge: Meta's Llama 3.2 3B and Google's Gemma 2 2B. While these transformer-based models are highly capable, their deployment as autonomous agents on phones, laptops, and IoT devices is severely constrained by memory bandwidth.

• Memory Footprint: Where a transformer-based 3B model typically requires heavy 4-bit quantization to run comfortably on a standard smartphone, LFMs operate with natively low resource overhead, preserving reasoning quality.
• Context Latency: In agentic workflows—where an agent must repeatedly read system prompts, tool definitions, and user histories—transformers suffer from severe "time-to-first-token" lag. LFMs bypass this bottleneck with highly efficient state-tracking.
• Throughput: Because LFMs maintain a constant-size state representation, throughput remains stable even during long-running agent loops, preventing the performance degradation common in long-context transformer sessions.

"We don't just need smaller models; we need architectures designed from the ground up to treat time and sequence as continuous variables. That is how we unlock true local intelligence."

Ramin Hasani, Co-founder & CEO of Liquid AI

Why Local Agents Are the Holy Grail

For developers and enterprise buyers, the transition from cloud-hosted APIs to local agents is not just about reducing latency—it is about privacy, reliability, and unit economics. An agent tasked with managing local files, monitoring desktop workflows, or controlling IoT devices cannot rely on a constant, high-bandwidth connection to a centralized cloud. It must run locally.

By releasing the LFM 2.5-2.6B models under an accessible open-weights framework on Hugging Face, Liquid AI is giving the developer community the raw materials to build non-transformer-based local agents. This release bypasses the token-metered toll booths of closed API providers, allowing developers to build agent loops that run indefinitely for zero marginal cost.

The Edge-Native Future

The release of these models signals a broader industry realization: the future of edge computing belongs to alternative architectures. While scale will continue to dominate the cloud, the edge demands efficiency that the transformer's attention mechanism simply cannot deliver. Liquid AI's latest release proves that small, alternative architectures can go toe-to-toe with the industry's heaviest hitters, setting the stage for a new generation of truly autonomous, local AI assistants.