Liquid AI Releases LFM2.5: A Compact AI Model Family For Real On Device Agents

2026-01-06 · Source: MarkTechPost · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems, Emerging Technologies & Innovation · Depth: Intermediate, short

Summary

Liquid AI has released LFM2.5, a new family of compact foundation models designed for on-device and edge deployments, building upon the LFM2 architecture. This family includes LFM2.5-1.2B-Base and LFM2.5-1.2B-Instruct, alongside specialized Japanese, vision language (LFM2.5-VL-1.6B), and audio language (LFM2.5-Audio-1.5B) variants. The models are available as open weights on Hugging Face and through the LEAP platform. The 1.2 billion parameter backbone's pretraining was extended from 10T to 28T tokens, with the instruct variant undergoing extensive fine-tuning and reinforcement learning. LFM2.5-1.2B-Instruct achieves 38.89 on GPQA and 44.35 on MMLU Pro, outperforming other 1B-class open models like Llama-3.2-1B Instruct and Gemma-3-1B IT on instruction following and function calling benchmarks.

Key takeaway

For CTOs and VPs of Engineering evaluating AI solutions for edge computing, LFM2.5 presents a compelling option due to its compact size and strong performance. Its specialized multimodal and language variants, including Japanese, vision, and audio, offer tailored capabilities for diverse on-device agent applications. You should consider integrating these open-weight models to enhance local processing capabilities and reduce reliance on cloud-based inference for latency-sensitive or privacy-critical workloads.

Key insights

LFM2.5 offers a compact, multimodal AI model family optimized for efficient on-device and edge deployments.

Principles

• Hybrid architectures enable fast, memory-efficient inference.
• Extended pretraining and multi-stage RL enhance instruction following.

Method

LFM2.5's training extends pretraining from 10T to 28T tokens, followed by supervised fine-tuning, preference alignment, and large-scale multi-stage reinforcement learning for instruction following and tool use.

In practice

• Deploy LFM2.5-VL-1.6B for document understanding on edge devices.
• Utilize LFM2.5-Audio-1.5B for real-time speech-to-speech agents.

Topics

• LFM2.5 Model Family
• On-Device AI
• Small Foundation Models
• Multimodal AI
• Benchmark Performance

Best for: CTO, VP of Engineering/Data, Director of AI/ML, AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Liquid AI's smallest model yet LFM2.5-230M beats models 4X its size at data extraction, can run 'anywhere' — Liquid AI's LFM2.5-230M demonstrates that highly efficient, small models can surpass larger ones for specific on-device data extraction and tool-use tasks.
• LFM2.5 — LFM2.5 offers a hybrid architecture for efficient, high-performance on-device AI with a tiny footprint.
• Seed1.8 Model Card: Towards Generalized Real-World Agency — Seed1.8 unifies LLM/VLM capabilities with multi-step agentic interaction for generalized real-world task execution.
• Liquid AI Releases LFM2.5-8B-A1B: An On-Device MoE Model With 8.3B Total and 1.5B Active Parameters — LFM2.5-8B-A1B is an on-device MoE model that significantly reduces hallucinations and enables complex tool-calling.
• IEEE Transactions on Artificial Intelligence, Volume 7, Issue 2, February 2026 — The volume showcases diverse AI advancements, from LLM applications in autonomous driving to specialized models for protein language processing.
• Gemma, DeepMind's Family of Open Models — Omar Sanseviero, Google DeepMind — Gemma 4 offers highly capable, open, and on-device LLMs with an Apache 2 license and multimodal multilingual support.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by MarkTechPost.