Model Releases
LFM2.5-24B-A2B: The New King of On-Device AI Agents
Liquid AI breaks the efficiency barrier with LFM2.5-24B-A2B, a 24B MoE model that delivers elite reasoning with only 2B active parameters for local execution.
13 октября 2026 г.
Model ReleaseLFM2.5-24B-A2B
Introduction: The Era of Local Intelligence
For developers building autonomous agents, the trade-off has always been clear: cloud-based models offer massive reasoning capabilities but sacrifice privacy and latency, while local models offer speed and security but often lack the 'brainpower' required for complex multi-step tasks. On October 13, 2026, Liquid AI fundamentally shifted this paradigm with the release of LFM2.5-24B-A2B.
LFM2.5-24B-A2B is not just another LLM; it is a purpose-built reasoning engine designed to live on your laptop or PC. As the largest member of the LFM2.5 family, it leverages a sophisticated Mixture-of-Experts (MoE) architecture to provide high-tier intelligence while maintaining a tiny active parameter footprint. This release marks a turning point for private, high-throughput agentic workflows.
- Released: October 13, 2026
- Primary Goal: Fast, private, on-device AI agents
- Architecture: Mixture-of-Experts (MoE)
- License: LFM Open License (Free for commercial use up to $10M revenue)
Architecture & Technical Specifications
The technical brilliance of LFM2.5-24B-A2B lies in its efficiency. While the model possesses 24 billion total parameters in its memory, it only activates approximately 2 billion parameters per token. This MoE approach allows the model to retain a vast knowledge base while executing inference at speeds typically reserved for much smaller models.
The training regime was equally massive, with the model being trained on 33 trillion pretraining tokens. This scale ensures that the reasoning capabilities are not just superficial but grounded in a deep understanding of complex patterns. Furthermore, the context window has been significantly expanded to support long-form reasoning and large document processing.
- Total Parameters: 24B
- Active Parameters per Token: 2B
- Context Window: 131K tokens
- Vocabulary Size: 128K tokens
- Training Data: 33T tokens
Performance Benchmarks: Outperforming the Giants
In the world of agentic workflows, efficiency is measured by how many tokens are required to reach a logical conclusion. LFM2.5-24B-A2B is a standout performer here, using 47% fewer reasoning tokens than the DeepSeek-R1-Distill-32B. This efficiency translates directly into roughly 2x higher agent throughput, making it ideal for rapid-fire tool use and decision-making.
When compared to larger competitors, the results are even more striking. On agentic and tool-use benchmarks, LFM2.5-24B-A2B outperforms Gemma 4 27B and remains highly competitive with Qwen3.6-35B-A3B—a model that possesses roughly 50% more total and active parameters. This 'punching above its weight' capability is the hallmark of the Liquid AI architecture.