Description
LiquidAI LFM2.5-2.6B is part of the LFM2.5 family, a series of hybrid models crafted for on-device deployment. Building on the LFM2 architecture, it features a 128K context window and agentic post-training, making it competitive with models four times its size in tool use, instruction following, and multi-step agentic tasks. The model is trained using agentic reinforcement learning, enhancing its compatibility with popular agentic harnesses.
Read More
The model boasts efficient inference capabilities, achieving 220 tokens per second on an Apple M5 Max and 113 tokens per second on an AMD Ryzen CPU, all within a memory footprint of under 2.5 GB. It is a general-purpose text-only model with 2.69 billion parameters, 30 layers, and a vocabulary size of 128,000. The model supports a wide range of languages, including English, Arabic, Chinese, French, and more.
LFM2.5-2.6B is available in various formats, such as the base model for fine-tuning, a quantized format for CPU inference, and an ONNX Runtime format for cross-platform deployment. It is optimized for fast inference on Apple Silicon using the MLX framework and supports speculative decoding for faster outputs.
The model is recommended for agentic workloads, tool use, data extraction, and long-context workflows, though it is not ideal for agentic coding and knowledge-heavy tasks. It supports function calling in agentic workflows, allowing for seamless integration into existing systems. Fine-tuning is encouraged to achieve optimal results for specific use cases.
LFM2.5-2.6B has been benchmarked against other sub-10B models, demonstrating competitive performance across various metrics. Its efficient architecture allows for fast CPU and GPU inference, making it suitable for a wide range of applications.
- CategoryAI Models & LLMs
- Pricing ModelOpen source
- DeploymentCloud, Self-hosted
- PlatformsWeb, macOS, Windows
LiquidAI LFM2.5-2.6B Highlights
Hybrid model for on-device deployment
128K context window
Agentic post-training
Efficient inference
2.69 billion parameters
Supports multiple languages
Available in various formats
Optimized for Apple Silicon
Getting Started with LiquidAI LFM2.5-2.6B
Access model page: Visit the Hugging Face website
Authenticate: Log in to your Hugging Face account
Configure: Set up the model for your specific use case
Send prompts: Use the model for inference or agentic workflows
LiquidAI LFM2.5-2.6B's Use Cases
- Agentic Workflows
- Data Extraction
- Tool Use
- Long-Context Tasks
- Cross-Platform Deployment
