LFM2.5-230M

LFM2.5-230M is a text-generation model optimized for on-device deployment, hosted on gigarouter as an OpenAI-compatible API. It combines a hybrid architecture of convolution and attention layers with 230 million parameters and was trained on 19 trillion tokens. The model supports a 32,768-token context window across ten languages: English, Arabic, Chinese, French, German, Italian, Japanese, Korean, Portuguese, and Spanish. It underwent a three-stage post-training pipeline—supervised fine-tuning with distillation from LFM2.5-350M, direct preference optimization, and multi-domain reinforcement learning—making it well-suited for agentic tasks such as tool use and structured data extraction.

Key strengths

• Fast on-device inference: 213 tokens/second on a Galaxy S25 Ultra and 42 tokens/second on a Raspberry Pi 5, with a small memory footprint.
• Native function-calling support using a ChatML-like template, enabling straightforward integration into agentic pipelines.
• Day-one support for multiple inference frameworks including llama.cpp, MLX, vLLM, SGLang, and ONNX.

Benchmark performance

The following table shows LFM2.5-230M scores on a subset of benchmarks from the official evaluation, comparing favorably against larger models in its class.

Deployment use cases

LFM2.5-230M has been deployed in real-world settings such as a skill-selection layer on a Unitree G1 humanoid robot running on an NVIDIA Jetson Orin, demonstrating its capability for decomposing natural-language instructions into multi-step tool calls. It is not recommended for reasoning-heavy workloads such as advanced math, code generation, or creative writing. Further details are available in the Liquid AI blog post.