Liquid AI Unveils LFM2.5-24B-A2B: The New Gold Standard for Local Agentic Reasoning

When compared to larger competitors, the results are even more striking. LFM2.5-24B-A2B outperforms the Gemma 4 27B model and remains highly competitive with the Qwen3.6-35B-A3B. Notably, the Qwen model has approximately 50% more total and active parameters, yet Liquid AI's model holds its own on critical agentic benchmarks. Furthermore, it offers a 50% lower time-to-first-token (TTFT) compared to its predecessor, the LFM2-24B-A2B, making it feel instantaneous to the end user.

• 2x higher agent throughput vs DeepSeek-R1-Distill-32B
• Outperforms Gemma 4 27B on agentic tasks
• Competitive with Qwen3.6-35B-A3B (despite 50% fewer parameters)
• 50% lower TTFT than LFM2-24B-A2B

API Pricing & Licensing

Liquid AI has adopted a developer-friendly approach to distribution. The model is released under the LFM Open License, which allows for free commercial use for organizations with up to $10M in annual revenue. This makes it an incredibly attractive option for startups and growing tech companies looking to integrate local reasoning without massive licensing overhead.

As this is an open-weights model designed for local deployment (via llama.cpp, Ollama, etc.), there is no centralized API cost for self-hosted versions. For those utilizing Liquid AI's hosted infrastructure, pricing details are subject to their specific cloud endpoints, but the primary value proposition is the zero-cost local inference capability.

• License: LFM Open License (Free for <$10M revenue)
• Deployment: Local (GGUF, MLX) or Cloud
• Primary Cost Driver: Local compute/hardware

Use Cases: Where LFM2.5-24B-A2B Shines

Because of its high throughput and low latency, this model is purpose-built for 'Agent Harnesses.' If you are building autonomous agents that need to browse the web, execute code, or manage local file systems, LFM2.5-24B-A2B provides the necessary reasoning without the cloud-latency penalty. It is ideal for private, on-device workflows where data security is paramount.

Beyond agents, the model is a powerhouse for local RAG (Retrieval-Augmented Generation) and coding assistance. The 131K context window allows developers to feed entire modules of code into the model for debugging or refactoring, all while keeping the data strictly on their local machine. The release of post-trained weights alongside the base model also makes it an excellent candidate for domain-specific fine-tuning.

• Autonomous AI Agents (Tool-use & Automation)
• Local RAG (Privacy-focused knowledge retrieval)
• On-device Coding Assistants
• Private Enterprise Workflows

Getting Started: Deployment & Customization

Getting started with LFM2.5-24B-A2B is straightforward for anyone familiar with the modern LLM ecosystem. The weights are available on Hugging Face under the 'LiquidAI/LFM2.5-24B-A2B' repository. To ensure compatibility across various hardware, Liquid AI has provided optimized GGUF and MLX versions.

Developers can immediately deploy the model using popular tools like llama.cpp, LM Studio, or Ollama. For those looking to push the model further, the availability of post-trained weights makes fine-tuning and customization highly accessible, allowing you to tailor the reasoning capabilities to your specific application or industry requirements.

• Hugging Face: LiquidAI/LFM2.5-24B-A2B
• Supported Formats: GGUF, MLX
• Compatible Tools: llama.cpp, LM Studio, Ollama

Conclusion

Liquid AI has successfully bridged the gap between massive-scale reasoning and local-scale efficiency. With LFM2.5-24B-A2B, the barrier to running high-performance, private, and autonomous agents on consumer hardware has effectively been removed. For developers building the next generation of AI-driven software, this model is a must-try.

• A breakthrough for local agentic workflows
• High efficiency via 2B active parameters
• Open weights for maximum flexibility

Sources