- LFM2.5-230M is a small, fast model for fine-tuning and deployment in agentic workflows, with fast inference on various devices including cloud GPUs and CPUs.
- It was pre-trained on 19T tokens and post-trained via supervised fine-tuning, direct preference optimization, and multi-domain reinforcement learning, balancing capabilities and adaptability.
- The model was tested on a humanoid robot for skill selection, turning natural-language instructions into structured plans using NVIDIA's SONIC framework.
- Benchmarks show LFM2.5-230M competes with or outperforms larger models in knowledge, instruction following, data extraction, and tool use tasks.
- It supports fast inference across ecosystems like llama.cpp, MLX, vLLM, and ONNX, with optimized CPU and GPU performance for low latency and high throughput.
- LFM2.5-230M is open-weight, available on Hugging Face, and designed for large-scale data extraction or lightweight on-device workloads, but not for reasoning-heavy tasks.