Andrii Bidochko

• Updated: January 28, 2026
• 6 min read

Liquid AI Unveils LFM‑2.5‑1‑2B: A 1.2 B‑Parameter On‑Device Reasoning Model

Liquid AI’s LFM‑2.5‑1‑2B is a 1.2 billion‑parameter reasoning model that runs fully on‑device, occupies less than 1 GB of storage, and delivers high‑quality structured reasoning for edge AI applications.

Liquid AI Unveils LFM‑2.5‑1‑2B: A Compact Reasoning Model Built for On‑Device AI

Imagine a language model that can solve math problems, plan tool calls, and generate step‑by‑step reasoning—all without ever touching a data‑center. Liquid AI’s newly released LFM‑2.5‑1‑2B makes that vision a reality, fitting comfortably on a modern smartphone while delivering performance that rivals many cloud‑hosted giants. For developers, startups, and enterprises hunting a compact language model that respects privacy and latency constraints, this release marks a turning point in the edge AI landscape.

What Is Liquid AI and the LFM‑2.5‑1‑2B Model?

A Brief Look at Liquid AI

Founded by a team of former research engineers, Liquid AI focuses on democratizing advanced language models for on‑device deployment. Their LFM‑2.5 family extends the original LFM2 architecture with deeper pre‑training and multi‑stage reinforcement learning, specifically engineered for low‑power hardware.

Key Technical Specs of LFM‑2.5‑1‑2B

• ≈ 1.17 B parameters (marketed as a 1.2 B‑parameter model)
• 16 transformer layers with 10 double‑gated LIV convolution blocks and 6 GQA blocks
• Training budget of 28 T tokens, context window of 32,768 tokens
• Vocabulary of 65,536 tokens covering eight major languages (English, Arabic, Chinese, French, German, Japanese, Korean, Spanish)
• Model size ≈ 900 MB (fits under 1 GB RAM on most smartphones)
• Optimized for “thinking” – produces explicit reasoning traces before final answers

These specifications enable the model to run on CPUs, NPUs, and even low‑end ARM cores, delivering up to 239 tokens/second on a desktop AMD CPU and 82 tokens/second on a Qualcomm mobile NPU. The model is available in GGUF, ONNX, and MLX formats, making it compatible with popular runtimes such as llama.cpp, vLLM, and FastFlowLM.

Why Compact Language Models Matter for On‑Device AI

Edge AI is no longer a niche; it’s a strategic imperative for privacy‑first applications, real‑time responsiveness, and cost‑effective scaling. A compact language model like LFM‑2.5‑1‑2B offers three decisive advantages:

1. Privacy by design: All inference happens locally, eliminating the need to transmit sensitive data to the cloud.
2. Latency reduction: Sub‑second response times are achievable because the model runs on the same device that captures the input.
3. Cost efficiency: No recurring cloud‑compute fees; developers can ship a single binary that works offline.

These benefits align perfectly with the goals of on‑device machine learning initiatives championed by forward‑thinking enterprises.

Performance Benchmarks & Real‑World Use Cases

Liquid AI evaluated LFM‑2.5‑1‑2B against a suite of reasoning and instruction benchmarks. The model consistently outperformed other 1 B‑class competitors, especially in math and tool‑use tasks.

These numbers translate into concrete use‑cases:

• Agentic workflows: The model’s “thinking” traces act as a planning brain for autonomous agents, enabling reliable tool calls and verification steps.
• Data extraction pipelines: Structured reasoning helps parse invoices, receipts, or legal documents on‑device without sending raw data to a server.
• Retrieval‑augmented generation (RAG): By generating intermediate reasoning, the model improves answer fidelity when combined with local vector stores.
• Mobile education apps: Real‑time math tutoring can now run offline, preserving student privacy while delivering instant feedback.

Developers building AI marketing agents can embed LFM‑2.5‑1‑2B as the core reasoning engine, while delegating knowledge‑heavy tasks to larger cloud models when needed.

Deployment Formats, Hosting Options, and Integration Paths

Liquid AI provides the model in three widely adopted formats, ensuring seamless integration with existing pipelines:

• GGUF: Optimized for llama.cpp and low‑memory environments.
• ONNX: Compatible with Microsoft’s ONNX Runtime for cross‑platform deployment.
• MLX: Tailored for Apple Silicon and other GPU‑accelerated devices.

For teams that prefer a no‑code approach, the model can be imported directly into the Workflow automation studio or the Web app editor on UBOS. Both tools let you drag‑and‑drop the model, configure input‑output schemas, and expose a REST endpoint in minutes.

If you need tighter control, the UBOS platform overview explains how to spin up a containerized micro‑service that serves the model via vLLM or FastFlowLM. Pricing is transparent through the UBOS pricing plans, which include a free tier for developers experimenting with edge models.

Ready to prototype instantly? Grab the UBOS templates for quick start and select the “AI Article Copywriter” template. Replace the default LLM with LFM‑2.5‑1‑2B, and you’ll have a fully offline copy‑writing assistant that can explain its reasoning step‑by‑step.

Availability, Roadmap, and How to Get Started

LFM‑2.5‑1‑2B is publicly available on Hugging Face under a permissive license. You can download the weights in any of the three formats mentioned above and begin inference within hours.

The roadmap promises two major upgrades in the next 12 months:

1. Multilingual expansion: Adding support for Hindi, Portuguese, and Russian while keeping the model size under 1 GB.
2. Hardware‑accelerated kernels: Native integration with Qualcomm Hexagon DSPs and Apple Neural Engine for sub‑50 ms latency on flagship phones.

Businesses interested in early access can join the UBOS partner program, which offers co‑marketing, dedicated support, and custom fine‑tuning services.

For startups looking to embed edge AI from day one, the UBOS for startups page outlines a fast‑track onboarding process, while SMBs can explore UBOS solutions for SMBs to leverage the model without heavy DevOps overhead.

Visual Overview

Figure: High‑level architecture of the LFM‑2.5‑1‑2B reasoning model.

For a deeper dive into the original announcement, see the original MarkTechPost article.

Conclusion: Why LFM‑2.5‑1‑2B Is a Game‑Changer for Edge AI

Liquid AI’s LFM‑2.5‑1‑2B delivers a rare combination of compact size, high‑quality reasoning, and on‑device friendliness**. Whether you are building a privacy‑first chatbot, a mobile tutoring app, or an autonomous agent that must verify its own steps, this model gives you the confidence to run sophisticated AI locally.

Ready to experiment? Visit the UBOS homepage to explore the full suite of tools that make deploying LFM‑2.5‑1‑2B effortless, from the latest AI news to hands‑on tutorials.

Take the next step: download the model, spin up a sandbox in the Web app editor, and watch your on‑device AI capabilities soar.

---

Andrii Bidochko

CTO UBOS

Andrii Bidochko is an AI entrepreneur and researcher focused on AI agents, reinforcement learning, and autonomous systems. He writes about the technologies shaping the future of machine intelligence, from frontier models and agent architectures to real-world AI applications.