Performance Metrics and Deployment Observations

Early user testing of LiquidAI’s LFM2.5-24B-A2B model suggests significant advantages in processing speed, particularly when deployed on high-end hardware. According to reports from users on Reddit, the model has demonstrated sustained speeds of approximately 95 tokens per second (tps) when running via llama.cpp using ROCm on an AMD W7900 workstation.

Some users have noted that the model’s throughput appears to be roughly double that of the Qwen 3.6 35B-13B model. In terms of memory requirements, the Q8 quantization of the model reportedly utilizes 30GB of VRAM when operating with a full 128k context window. While these performance figures are based on individual user configurations and lack independent, third-party verification, they indicate the model is being utilized for high-speed, recursive, and mechanical tasks within agentic workflows.

Integration in Agentic Workflows

The LFM2.5-24B-A2B is being tested as a “workhorse” model within multi-model architectures, such as the “Pi Agent” framework. In these setups, the model is often used alongside larger, more computationally expensive models to handle specific sub-tasks.

Early feedback from the developer community suggests the model is competent in executing tool calls and maintaining stability during long-running tasks. However, some users have raised concerns regarding its accuracy in specialized domains. For instance, reports indicate the model may struggle with complex cybersecurity reporting and certain aspects of world knowledge, occasionally providing inaccurate definitions for non-encyclopedic queries.

Technical Implementation and Limitations

While the model shows promise in speed and task reliability, some users have highlighted potential friction points in the deployment process. Issues regarding chat templates—which can impact a model’s performance relative to its original benchmarks—have been noted by community members during the release phase.

Furthermore, discussions among users suggest that while the model excels in speed and mechanical task execution, it may not yet match the writing style or “agentic” capabilities of competitors like Google’s Gemma series. As with many new model releases, the practical utility of the LFM2.5 series in complex “council-mode” workflows remains subject to ongoing user testing.