Deploy local agents everywhere with LFM2.5-2.6B

The landscape of edge computing is shifting as Liquid AI unveils LFM2.5-2.6B, a compact yet remarkably potent model engineered to bring sophisticated agentic capabilities directly to consumer hardware. By balancing a lightweight footprint with high-level reasoning, this model enables developers to bypass cloud-based inference costs and latency, ensuring that sensitive data remains localized on the user’s device—whether it is a smartphone or a high-performance laptop.

Redefining Edge Intelligence

LFM2.5-2.6B is not just another small language model; it is a specialized engine designed for tool calling, multi-step workflows, and complex instruction following. Despite its diminutive size, it consistently punches above its weight, delivering performance metrics that rival models up to four times its parameter count.

By utilizing agentic reinforcement learning, the model has been trained within the most widely used agentic harnesses. This ensures that when developers integrate the model into their own workflows, it is already optimized for compatibility and reliability.

"LFM2.5-2.6B is built to power capable agents entirely on-device, enabling developers to scale usage without the burden of a cloud inference bill while keeping data private."

The Architecture of a Reliable Agent

The development of LFM2.5-2.6B involved a rigorous, multi-stage training pipeline designed to transform a base model—pre-trained on approximately 34 trillion tokens with a context window extended to 128K—into a highly effective agent. The post-training process consists of four critical phases:

1. Supervised Fine-Tuning (SFT): Two rounds of intensive training focused on agentic data, including web search, tool utilization, and harness trajectories. 2. Teacher Specialization: The creation of domain-specific teachers (math, code, and tools) to provide expert-level guidance. 3. Multi-Domain On-Policy Distillation (MOPD): Consolidating the knowledge of these specialist teachers into a single, unified student model. 4. Agentic Reinforcement Learning (Agentic RL): A sophisticated loop where the model learns to navigate multi-turn tasks, system prompts, and diverse tool environments.

The RL pipeline is particularly innovative, separating the Training Engine (model optimization), the Rollout Engine (action generation), and the Sandbox Service (execution). By employing a Harness Proxy, the team can treat existing agentic frameworks as "black boxes," allowing for seamless integration without modifying the underlying infrastructure while capturing the granular token-level trajectories necessary for validation.

Benchmark Performance: Small Model, Big Results

When pitted against larger competitors, LFM2.5-2.6B demonstrates exceptional versatility. It frequently outperforms models like the Gemma-4-E2B (5.1B) and Gemma-4-E4B (8B) across various benchmarks, including IFBench (Instruction Following) and ToolSandbox.

• Instruction Following: The model leads its class, proving superior in adhering to complex user prompts.
• Tool Use: It dominates most benchmarks, with only the significantly larger Qwen3.5-9B showing a marginal edge in specific tests like BFCLv4.
• Agentic Tasks: It remains neck-and-neck with larger models, proving that parameter count is not the sole determinant of agentic success.

While larger models still maintain a lead in pure coding tasks, LFM2.5-2.6B is the clear choice for developers prioritizing instruction-following and tool-based automation on resource-constrained hardware.

Unmatched Inference Speed

Efficiency is the hallmark of the LFM2 architecture. The model is designed for immediate deployment across the modern inference ecosystem, including llama.cpp, MLX, vLLM, SGLang, and ONNX.

• CPU Performance: On an Apple M5 Max, the model achieves a blistering 220 tokens/s, while an AMD Ryzen AI Max+ 395 handles it at 113 tokens/s. Even at a modest 30 tokens/s, the model remains highly viable for mobile deployment.
• GPU Throughput: In high-concurrency environments, the model reaches nearly 15,000 output tokens per second, translating to roughly 1.3 billion tokens per day on a single H100 GPU.

Getting Started

For developers looking to integrate LFM2.5-2.6B, the process is streamlined via the standard Hugging Face transformers library. By loading the model with device_map="auto", developers can immediately begin building local agents.

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "LiquidAI/LFM2.5-2.6B"
model = AutoModelForCausalLM.from_pretrained(model_id, device_map="auto", dtype="bfloat16")
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Example usage for agentic tasks
prompt = "What is C. elegans?"
# ... (Apply chat template and generate)

Conclusion

With the release of LFM2.5-2.6B and its base variant, Liquid AI is delivering on its promise of making AI accessible and functional anywhere. Whether you are building a research agent, a local automation tool, or a privacy-first assistant, this model provides the necessary foundation.

Developers are encouraged to explore the WebGPU demo to see the model in action directly in the browser, or to dive into the documentation to integrate it with existing harnesses like OpenClaw or Hermes Agent. The era of the truly local, capable agent has arrived.