Features
LFM2.5-1.2B-Instruct with best-in-class instruction following
LFM2.5-230M model for ultra-small edge devices
LFM2.5-VL-1.6B vision-language with multi-image comprehension
LFM2.5-VL-450M structured visual intelligence model
LFM2.5-Audio-1.5B native audio input/output with 8x faster detokenizer
LFM2.5-1.2B-JP Japanese-optimized chat model
LFM2.5-8B-A1B mixture-of-experts model (8B total, 1B active)
LFM2.5 Retrievers for fast multilingual search on edge
Open weights on Hugging Face and LEAP
Commercial use free under $10M annual revenue
Hybrid architecture for fast CPU inference with low memory
Reinforcement learning post-training pipeline
Supports fine-tuning for domain-specific assistants
Deployment via LEAP, llama.cpp, MLX, vLLM, ONNX
NPU optimization by AMD and Nexa AI
Local model execution for offline inference
Connect online APIs: OpenAI (ChatGPT), Anthropic (Claude), Google (Gemini), Meta (Llama), Mistral AI, Alibaba (Qwen),
123+ HuggingFace models available
Memory feature (coming soon) for context persistence
Minimalist UI with dark mode
Open source under MIT license on GitHub
Cross-platform desktop app (Mac, Windows, Linux)
Version 0.8.3 with bug fixes and performance improvements
