LFM2.5

Key Features

• •LFM2.5-1.2B-Instruct model delivers excellent instruction following and tool use capabilities out of the box, trained with supervised fine-tuning, preference alignment, and large-scale multi-stage reinforcement learning for reliable edge agentic AI.
• •LFM2.5-VL-1.6B vision-language model provides multi-image comprehension and multilingual vision understanding across seven languages with stronger instruction-following performance for multimodal on-edge applications requiring visual intelligence.
• •LFM2.5-Audio-1.5B processes audio natively without pipelined approaches, eliminating information barriers between components and reducing latency with a custom LFM-based audio detokenizer that's 8x faster than its predecessor on mobile CPUs.
• •LFM2.5-1.2B-JP model offers specialized optimization for Japanese language with cultural and linguistic nuance, achieving state-of-the-art performance on Japanese knowledge and instruction-following benchmarks for developers building Japanese applications.
• •Extensive framework support includes day-zero compatibility with llama.cpp for CPU inference, MLX for Apple Silicon, vLLM for GPU serving, ONNX for cross-platform deployment, and LEAP for iOS/Android deployment as easily as cloud API calls.
• •Optimized quantization includes INT4 quantization-aware training for audio models enabling seamless low-precision deployment with minimal quality loss, plus GGUF checkpoints for efficient deployment on any hardware with optimized quantization.
• •Low memory profile enables extremely fast inference on CPUs compared to similar-sized models, with benchmarks showing significantly reduced memory usage while maintaining competitive performance across knowledge and instruction following tasks.
• •Partner-optimized NPU performance through collaborations with AMD and Nexa AI delivers highly efficient on-device inference on Qualcomm Snapdragon platforms and AMD Ryzen AI hardware for specialized edge deployment scenarios.

Use Cases

• •Local copilots and productivity workflows where data privacy is paramount, enabling AI assistance without sending sensitive information to cloud servers while maintaining fast response times and offline functionality.
• •In-car assistants that process voice commands and visual information in real-time during driving scenarios, providing navigation help, vehicle control, and information retrieval without internet dependency for safety-critical applications.
• •Japanese-language applications requiring cultural and linguistic nuance, where the specialized LFM2.5-JP model handles language-specific knowledge and instruction-following for customer service, content creation, and educational tools in Japanese markets.
• •Multimodal edge devices that combine vision and language understanding, such as smart cameras that interpret scenes in multiple languages or retail systems that analyze visual merchandise with textual instructions for inventory management.
• •Voice-first IoT interfaces for smart home devices, industrial equipment, and mobile applications where low-latency audio processing enables natural conversations without cloud round-trip delays while preserving user privacy locally.
• •Embedded systems in vehicles, mobile devices, and constrained hardware where memory and computational resources are limited but AI capabilities are needed for real-time decision making and intelligent user interactions without external connectivity.

Who is this for?

Developers building edge AI applications for mobile, embedded, and IoT devices; enterprises deploying private AI solutions on-premises; automotive companies implementing in-vehicle assistants; IoT manufacturers adding intelligent capabilities to constrained hardware; Japanese market developers requiring culturally-aware AI; organizations needing multimodal AI (vision, audio, text) that operates without cloud dependency.