Run fast, private agents on-device with LFM2.5-24B-A2B

LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents

LiquidAI has released LFM2.5-24B-A2B, a new open-weight model engineered for running complex AI agents directly on consumer hardware. The model's design focuses on efficiency, using a Mixture-of-Experts (MoE) architecture with 24 billion total parameters but only 2 billion active during inference. This approach enables fast, private, and continuous agentic operations on devices like laptops, sidestepping the need for constant cloud connectivity and API calls.

Technical Profile and Performance

LFM2.5-24B-A2B is an evolution of its predecessor, expanding the context window to 131K tokens and doubling the vocabulary to 128K for improved multilingual support. Built as a reasoning-first model, it explicitly generates a chain of thought before providing a final answer. The training regimen included a scaled-up pre-training phase on approximately 33 trillion tokens, followed by a sophisticated five-stage post-training process that used agentic reinforcement learning inside harnesses like OpenClaw and Hermes Agent.

• Total Parameters: 24B
• Active Parameters: 2B (top-4 routing from 64 experts)
• Context Window: 131K tokens
• Inference Speed (CPU): 184 tok/s on Apple M5 Max
• Inference Speed (GPU): 21K+ output tok/s on a single NVIDIA B300

Ecosystem Impact and Competitive Landscape

Benchmark results position LFM2.5-24B-A2B as a strong competitor against larger MoE models like Qwen3.6-35B-A3B, particularly in agentic and tool-use tasks. It leads or matches on benchmarks such as LiveClaw, ToolSandbox, and AutomationBench, where it scored more than three times higher than the next-best model. While it excels at instruction following and tool use, the source notes that larger models maintain an advantage on knowledge-heavy and complex coding benchmarks. For developers, this positions the model as a powerful foundation for on-device applications that require reasoning and tool interaction, with the recommendation to pair it with retrieval systems for knowledge-intensive workloads.

LiquidAI's focus on minimizing active parameters while scaling total parameters signals a strategic push towards decoupling advanced AI reasoning from data center dependency, making persistent, on-device agents a more viable commercial reality.