Liquid AI Releases LFM2.5-24B-A2B for Local AI Workflows

Liquid AI has launched an open-weight 24 billion parameter model called LFM2.5-24B-A2B, enabling local AI agents to run on laptops and other hardware. This model expands the LFM2.5 family, which ranges between 230 million and 24 billion parameters. The new release includes an expanded context window from 32,768 to 131,072 tokens, an expanded vocabulary of 128,000 tokens for improved tokenization of non-Latin languages, and additional training focused on reasoning, mathematics, coding, instruction following, and multilingual text.

The model achieved higher scores in Liquid AI's benchmarks for agent workflows, knowledge, coding, mathematics, and instruction following compared to some other models. However, Qwen scored higher in some knowledge-focused tests. Liquid AI recommends combining the model with retrieval systems or web search for dependable access to extensive factual information.

IT teams should consider hardware capacity and test their chosen model format and context length on intended devices. Local inference may keep model inputs and outputs on company hardware, but an agent connected to web search or cloud applications may still transmit data outside the device. Liquid AI suggests IT teams restrict tool access, protect credentials, and require approval before consequential actions.

Licensing may affect deployment costs, with companies having annual revenue exceeding $10 million needing a separate commercial license.