Revolutionizing Local AI Agent Performance
Liquid AI recently announced the release of LFM2.5-24B-A2B, an innovative open-weight reasoning model. Positioned as a significant advancement for AI agents operating on laptops and personal computers, this model is the most capable and final iteration within the LFM2.5 series. Its design specifically emphasizes multi-step tool interaction rather than broad, open-ended conversational tasks.
Core Specifications and Accessibility
LFM2.5-24B-A2B operates as a Mixture-of-Experts (MoE) model, encompassing 24 billion total parameters while actively engaging approximately 2 billion parameters per token. A standout feature is its extensive context window, supporting up to 131,072 tokens. The model's base and post-trained weights are readily available for download on Hugging Face, enabling developers to integrate this technology into their projects.
Performance Benchmarks and Strategic Focus
Internal evaluations by Liquid AI indicate that LFM2.5-24B-A2B either leads or performs comparably to larger MoE models across various agentic and tool-use benchmarks. Notably, it demonstrates exceptional aptitude in automation tasks, achieving a score of 27.69 on AutomationBench, significantly surpassing competitors like Qwen3.6-35B-A3B. While excelling in instruction following and tool use, the model exhibits lesser proficiency in general knowledge and coding challenges, suggesting its optimal deployment involves pairing it with retrieval mechanisms for comprehensive tasks.
Liquid AI's comparative analysis against models such as Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning highlights specific strengths:
- Leading Performance: Dominant scores on AutomationBench, ToolSandbox, LiveClaw, and τ³-Bench Banking.
- Strong Instruction Following: Outperforms rivals in IFEval, IFBench, and Multi-IF.
- Areas for Enhancement: Larger models maintained leads in GPQA Diamond and SWE-Bench Verified, indicating that LFM2.5-24B-A2B is primarily optimized for its core agentic reasoning function.
Architectural Insights and Training Methodology
The model’s architecture maintains LFM2's hybrid 40-layer design, incorporating both double-gated short convolution blocks and grouped query attention blocks. Each MoE block features 64 experts with a top-4 routing mechanism. Key improvements from its predecessor include a substantial expansion of the context window and a doubled vocabulary size, enhancing its ability to process non-Latin languages. The extensive training regimen involved pre-training with 33 trillion tokens and a multi-stage post-training process, including supervised fine-tuning, teacher specialization, and agentic reinforcement learning within various agentic frameworks. Liquid AI also disclosed and addressed instances of 'reward hacking' during training, ensuring subsequent evaluations were conducted in secure, isolated environments.
Hardware Compatibility and Integration
LFM2.5-24B-A2B is optimized for efficient on-device execution. It achieves decoding speeds of 184 tokens per second on an Apple M5 Max and 114 tokens per second on an AMD Ryzen AI Max+ 395, utilizing Q4_K_M quantization. For server-side applications, it can process 21,000 output tokens per second on a single NVIDIA B300. The model supports a wide array of popular AI frameworks and libraries from day one, including Transformers, vLLM, SGLang, llama.cpp, MLX, LM Studio, and ONNX.
Licensing and Commercial Use
Liquid AI provides API access through Liquid Console, with pricing set at $0.10 per million input tokens and $1.00 per million output tokens. The model's weights are released under the LFM Open License v1.0, an Apache 2.0-derived license. This license permits free commercial use for companies with annual revenues below $10 million, fostering broader adoption and innovation.
This article is a rewritten summary based on publicly available reporting. For the original story, visit the source.
Source: MarkTechPost
