fine-tuning

Liquid AI releases a 230M model optimized for phones, Raspberry Pi, and robots

3 months ago

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.