Deploy local agents everywhere with LFM2.5-2.6B

LiquidAI Launches LFM2.5-2.6B to Power On-Device Agents

LiquidAI has released LFM2.5-2.6B, a compact 2.6 billion parameter model engineered to run sophisticated AI agents directly on consumer hardware. The model supports complex operations like tool calling and multi-step workflows while maintaining a small memory footprint of under 2.5 GB. This design allows developers to deploy applications on devices ranging from laptops to phones, ensuring user data remains private and eliminating the recurring costs associated with cloud-based inference.

Technical Architecture and Performance

The model's capabilities are the result of a multi-stage training process built upon a base pre-trained on approximately 34 trillion tokens. LiquidAI employed a specialized four-stage post-training regimen to optimize the model for agentic tasks. This included supervised fine-tuning on agentic data, domain-specific teacher specialization, and multi-domain on-policy distillation. The final stage, Agentic Reinforcement Learning (Agentic RL), trained the model within common agent harnesses to refine its real-world performance.

• Pre-training: ~34T tokens with a 128K context window.
• Training Stages: Supervised fine-tuning, teacher specialization, multi-domain distillation, and Agentic RL.
• CPU Inference Speed: 220 tokens/second on an Apple M5 Max and 113 tokens/second on an AMD Ryzen AI Max+ 395.
• Compatibility: Supported by llama.cpp, MLX, vLLM, SGLang, and ONNX.

Market Impact and Competitive Standing

According to benchmark results, LFM2.5-2.6B is highly competitive with models up to four times its size, including gemma-4 and Qwen3.5 variants. It demonstrates strong performance in instruction following and tool use benchmarks, often outperforming its larger counterparts. While it remains competitive on general agentic tasks, the provided data indicates that larger models still hold an advantage in specialized coding tasks. The release signifies a growing trend toward powerful, efficient edge AI that can operate independently of large-scale data centers, potentially lowering the barrier to entry for developing and deploying agent-based applications.

LiquidAI's focus on on-device performance with LFM2.5-2.6B signals a growing market demand for efficient, private, and cost-effective agentic AI that operates outside the centralized cloud infrastructure.