LFM2.5

Key Features

• •LFM2.5-1.2B-Instruct Model: A general-purpose, instruction-tuned variant trained with supervised fine-tuning and multi-stage reinforcement learning. It delivers excellent instruction following and tool use capabilities out of the box, achieving top performance on knowledge, math, and instruction benchmarks while maintaining blazing inference speed on CPUs.
• •LFM2.5-VL-1.6B Vision-Language Model: A refreshed multimodal model built on an updated backbone for stronger real-world performance. It offers improved multi-image comprehension and multilingual vision understanding in seven languages, making it ideal for on-edge applications requiring visual and textual intelligence.
• •LFM2.5-Audio-1.5B Audio-Language Model: A native model that accepts both speech and text as input and output, processing audio natively to eliminate pipeline latency. Its custom LFM-based audio detokenizer is 8x faster than the previous version on mobile CPUs, enabling real-time voice interactions on constrained hardware.
• •LFM2.5-1.2B-JP Japanese Model: A chat model specifically optimized for Japanese language and cultural nuance. It pushes state-of-the-art performance on Japanese knowledge and instruction-following benchmarks, ideal for developers building linguistically accurate applications.
• •Open-Weight and Framework Support: All models are open-weight, available for download and fine-tuning without restrictions. They feature day-zero support across inference frameworks like llama.cpp for CPU, MLX for Apple Silicon, vLLM for GPU, and ONNX for cross-platform deployment.
• •Extreme Inference Efficiency: The models are optimized for low memory profile and high speed on edge hardware. Benchmarks show extremely fast inference on CPUs and NPUs, with quantized versions (e.g., INT4) maintaining high quality while drastically reducing memory footprint.
• •Comprehensive Model Family: Includes Base models for customization, Instruct models for general use, and specialized variants for vision, audio, and Japanese. This provides a single, cohesive architecture covering diverse on-device use cases from text to multimodal applications.
• •NPU Optimization via Partnerships: Through partnerships with AMD and Nexa AI, LFM2.5 models are optimized for Neural Processing Units (NPUs), enabling highly efficient on-device inference on a wide range of hardware from mobile phones to IoT devices.

Use Cases

• •Local Copilots and Productivity Workflows: Deploy the LFM2.5-Instruct model on laptops or mobile devices to create AI assistants that work offline. This provides private, fast text generation and task automation without cloud dependency, enhancing user productivity with always-available intelligence.
• •In-Car Voice Assistants: Utilize the LFM2.5-Audio model in vehicles for real-time, voice-driven interactions. Its native audio processing and 8x speed improvement enable low-latency speech recognition and synthesis, creating responsive and safe hands-free control systems.
• •Multimodal IoT and Embedded Vision: Implement the LFM2.5-VL model on cameras or sensors for smart vision applications. Its multi-image comprehension and multilingual capabilities allow for on-device analysis like object detection or scene understanding in diverse environments without sending data to the cloud.
• •Japanese-Language Applications: Build chatbots, content tools, or customer service agents for the Japanese market using the LFM2.5-JP model. Its optimization for linguistic and cultural nuance ensures accurate and context-aware interactions in Japanese.
• •Edge AI Prototyping and Customization: Researchers and developers can use the LFM2.5-Base model for fine-tuning on proprietary datasets. This enables creation of domain-specific assistants or experimentation with novel post-training approaches for specialized edge AI solutions.
• •Cross-Platform AI Deployment: Deploy a single model across diverse hardware—from Apple Silicon Macs to Android phones to AMD NPUs—using supported frameworks like MLX, llama.cpp, and ONNX. This simplifies development and ensures consistent performance for applications targeting multiple device ecosystems.

Who is this for?

Developers and engineers building edge AI applications for consumer and enterprise products. Enterprises seeking custom, on-device AI solutions for data privacy and low latency. Product teams working on vehicles, mobile devices (iOS/Android), laptops, and IoT/embedded systems. Researchers and hobbyists experimenting with open-weight model fine-tuning and deployment. Companies partnering with hardware vendors like AMD and Qualcomm for NPU-optimized AI inference.