Liquid AI Releases LFM2.5-24B-A2B: Fast On-Device Agent Model
Liquid AI has released LFM2.5-24B-A2B on Hugging Face, combining 24B parameters with 2B active activation for fast, private on-device agents.
👁 113 reads
On August 5, 2025, Liquid AI published the LFM2.5-24B-A2B model on Hugging Face - an architecture engineered specifically for fast, private AI agents running directly on end-user hardware. It is one of the few models in its class to combine a large nominal parameter count with an extremely low activation cost.
24 Billion on Paper, 2 Billion in Practice
The numbers in the designation are quite deliberate. LFM2.5-24B-A2B is a model featuring 24 billion parameters in total, but its Mixture of Experts (MoE) mechanism ensures that only 2 billion of them are activated during any given forward pass. While this setup is familiar from architectures like Mistral’s Mixtral or Google Gemini, Liquid AI applies it to a context that has rarely been prioritized: running agents locally, fully independent of cloud infrastructure.
What does this mean in everyday deployment? The model can run on consumer hardware or edge computing devices without requiring data center resources. The active footprint of 2B parameters means memory bandwidth and compute demands stay close to compact 2-3B models, yet the system retains a significantly richer knowledge representation stored across the full 24 billion parameters - dynamically routed based on query context.
For developers architecting agentic workflows, this offers a substantial shift. A typical AI agent executes dozens or even hundreds of model calls to complete a single user task - inspecting state, planning steps, calling tools, and validating outputs. Each call made to a cloud model adds hundreds of milliseconds of latency alongside network overhead. An on-device model eliminates external data transit and drastically shortens that execution loop.
Privacy as an Architectural Foundation
Liquid AI directly positions LFM2.5-24B-A2B for environments where data privacy is an absolute requirement rather than an afterthought. The on-device model market is accelerating under two primary pressures: regulatory compliance (such as GDPR across Europe, along with strict sector-specific rules in finance and healthcare) and growing reluctance among organizations to route proprietary data through third-party servers.
Across regulated enterprise environments, teams implementing AI in human resources, legal analysis, or clinical processes routinely face an uncomfortable tradeoff: rely on massive cloud APIs while managing data compliance risks, or fall back on smaller, less capable local models. LFM2.5-24B-A2B aims to carve out a middle path - powerful enough to handle complex reasoning, yet compact enough to execute on local silicon.
This is not the sole initiative addressing the local footprint. Apple Intelligence relies on an on-device-first strategy with selective cloud escalation. Microsoft continues to iterate on Phi models for Copilot+ PCs, while Qualcomm and NVIDIA supply specialized SDKs optimized for local neural processing. Liquid AI enters this domain with a design tuned specifically for multi-step agentic routines rather than generic benchmarks.
LFM - Moving Beyond the Standard Transformer
The underlying mechanics of the Liquid Foundation Models (LFM) family set it apart. Founded by researchers from MIT, Liquid AI builds on Liquid Neural Networks - systems inspired by biological nervous architectures where connection dynamics adapt in real time to incoming input data. This fundamentally diverges from the standard Transformer backbones driving models like GPT, Claude, Gemini, or Llama.
Earlier LFM releases demonstrated that this structure handles extended contexts and sequential data processing with notable efficiency - precisely the characteristics demanded by an agent orchestrating multi-phase workflows. The model card on Hugging Face presents LFM2.5-24B-A2B as the next stage in this lineage, integrating MoE routing to expand total capacity while keeping per-token inference overhead minimal.
LFM2.5 represents an evolution over the prior LFM2 generation. Although Liquid AI maintains a leaner profile than giants like OpenAI, Anthropic, or Mistral, reception on Hugging Face - drawing over 2,910 likes and 351 downloads within days of release - highlights steady interest from the open-source engineering community.
Google · Your sourcesQuantslant higher in your Google - one clickAdd to preferred sources →What It Means for AI Application Builders
The rollout of LFM2.5-24B-A2B aligns with a broader shift that accelerated through 2024 and 2025: decentralized inference. While industry focus long centered on ever-larger frontier cloud clusters, an increasing number of hardware providers and software labs envision a hybrid split: massive cloud infrastructures reserved for model training and rare high-complexity queries, paired with efficient, specialized on-device weights handling continuous agent operations.
For software engineers building agent platforms, this opens a practical alternative: instead of tethering an entire architecture to external API endpoints and incurring per-call fees, a local model can serve as the primary execution engine. With LFM2.5-24B-A2B available on Hugging Face, developers can download the weights and self-host on their own hardware - free from token bills, rate caps, and unpredictable API policy shifts.
The central question that remains is how LFM2.5-24B-A2B holds up across standardized agentic benchmarks relative to established alternatives like Mistral Small, Phi-3.5, Gemma, or Qwen. Liquid AI has not yet released comprehensive comparative evaluations, and standardized figures will ultimately determine how widely this architecture is adopted beyond specialized research setups.
