Deploying Local Agents Everywhere with LFM2.5-2.6B: A Comprehensive AI Solution for Businesses and Data Analysis Tools

A new model has been released that enables the deployment of local agents everywhere, making it easier to integrate artificial intelligence into various applications. The LiquidAI team has developed LFM2.5-2.6B, a pre-trained language model designed specifically for edge devices. This AI solution is ideal for businesses and organizations looking to leverage data analysis tools in their operations.

The key feature of LFM2.5-2.6B is its ability to perform competitively with larger models on various tasks such as tool use, instruction following, and multi-step agentic tasks. In fact, it has been shown that this model can outperform models up to four times its size in certain areas. This makes it an attractive option for developers who need a reliable AI solution without the high computational requirements.

The development of LFM2.5-2.6B involved training the model on approximately 34 trillion tokens, with a mid-training phase that extended the context window to 128 kilobytes. The post-training process consisted of four stages: supervised fine-tuning, teacher specialization, multi-domain on-policy distillation, and agentic reinforcement learning. Each stage was designed to improve the model’s performance in specific areas such as tool use, web search, and harness trajectories.

The agentic reinforcement learning pipeline is a crucial component of LFM2.5-2.6B’s architecture. It separates model optimization, inference, and environment execution into distinct components, allowing for more efficient training and deployment. The Training Engine optimizes the model, while the Rollout Engine generates actions using the latest policy. The RL framework orchestrates the training loop by launching rollouts, collecting trajectories and rewards, and updating the model.

The Sandbox Service is another essential part of LFM2.5-2.6B’s architecture. It hosts the agent (e.g., OpenClaw or Hermes Agent) and coordinates interactions with the task environment. The Harness Proxy allows developers to treat agentic harnesses as black boxes without modification, while capturing token-level trajectories needed for reconstructing and validating RL training samples.

Benchmark results have shown that LFM2.5-2.6B performs exceptionally well on various tasks such as instruction following, tool use, and agentic tasks. It tops every instruction-following benchmark and beats the rest in most cases. On agentic tasks, it outperforms both Gemma models and stays even with Qwens. Additionally, LFM2.5-2.6B leads on knowledge retrieval and stays close on math problems.

Inference speed is another area where LFM2.5-2.6B excels. It can process 220 tokens per second on an Apple M5 Max CPU and 113 tokens per second on an AMD Ryzen CPU, all within under 2.5 gigabytes of memory. This makes it suitable for running capable agents even on a phone. On GPU inference, LFM2.5-2.6B reaches almost 15 kilotokens per second at high concurrency, roughly 1.3 billion tokens per day on a single H100.

The model is designed to be used in various applications and can be integrated with existing infrastructure using the latest version of transformers (compatible with transformers>=5.0.0). The installation process involves running pip install -U transformers followed by loading and running the model using AutoModelForCausalLM and AutoTokenizer from the transformers library.

The LFM2.5-2.6B demo is available for testing, showcasing its capabilities in a research agent that helps users generate summaries of specific questions. The model can be downloaded on Hugging Face today, along with the LFM2.5-2.6B-Base variant. With this release, LiquidAI aims to deliver AI solutions that run anywhere, making it easier for developers and businesses to integrate data analysis tools into their operations.