Liquid AI Releases LFM2.5

Liquid AI released the on-device agent model LFM2.5-2.6B with 2.69B parameters and LFM2.5-2.6B-Base for fine-tuning. Both models are available on Hugging Face.

These models aim to run agents on mobile phones and CPUs instead of cloud APIs, and were post-trained with a focus on tool calling and multi-step task execution. Local execution is designed to reduce token costs and network latency, ensuring data does not leave the device.

Key specifications are as follows:

• 128K token context length
• 128K vocabulary and support for 16 languages
• Composition of 22 gated convolution blocks and 8 GQA attention blocks out of 30 layers
• 2,048 hidden dimensions, 32 attention heads, and 8 KV heads
• Approximately 34 trillion tokens of pre-training
• Support for 16 languages including English, Chinese, Japanese, and Korean

Liquid AI explained that the models were further trained using agentic reinforcement learning with tool lists and interaction patterns from specific agent harnesses. The hybrid architecture, centered on convolution blocks, is designed to reduce KV cache memory and inference traffic, making it suitable for mobile and CPU environments.