Liquid AI this week published details on Hackernews about its latest model, LFM2.5-24B-A2B, a 24-billion-parameter language model built from the ground up to run agentic tasks — multi-step, tool-using AI workflows — entirely on local hardware rather than through remote servers. The model belongs to Liquid's second-generation LFM (Liquid Foundation Model) series and is optimized specifically for the "A2B" (Agent-to-Benchmark) use case, targeting fast inference on consumer-grade devices including laptops and on-device platforms.
The distinguishing engineering claim is the model's active parameter footprint: despite having 24 billion total parameters, the architecture activates only roughly 2 billion parameters per forward pass, a mixture-of-experts-style efficiency that allows it to run substantially faster and with lower memory pressure than a dense 24B model would require. Liquid states the model achieves competitive scores on standard agentic benchmarks while maintaining the throughput needed for real-time agent loops — where a model must reason, call tools, evaluate results, and iterate, often dozens of times per task.
Liquid AI positions the release explicitly against the assumption that capable agentic AI requires cloud infrastructure. The company's published figures indicate the model is quantizable to formats that fit within the VRAM budgets of current high-end consumer GPUs, and the firm has released weights for local deployment.
What the mainstream technology press is unlikely to dwell on is the infrastructure dependency profile this class of model breaks. Every cloud-dependent AI agent — whether it is summarizing documents, managing a calendar, or operating a home automation system — requires a continuous, low-latency internet connection and the continued commercial operation of a remote API provider. A capable 24-billion-parameter model that fits on local hardware and runs agent loops without phoning home changes that dependency structure in a concrete way. For households and small organizations that have invested in offline-capable computing as part of broader resilience planning, this architecture represents a meaningful capability threshold: agentic reasoning that does not degrade or disappear when connectivity is interrupted, rate-limited, or priced out of reach. Our overview of local AI hardware considerations covers the current device landscape for running models in this parameter range.
Liquid AI is a Boston-based research company founded in 2023, and the LFM2.5 series represents its most capable public model release to date. The company did not announce pricing for commercial API access to the 24B variant in the initial release post, focusing instead on the on-device deployment pathway.
