Liquid AI has released LFM2.5-24B-A2B, a 24-billion-parameter reasoning model that, in a gesture of restraint humanity could stand to study, only uses 2.45 billion of those parameters at any given moment. It runs locally, thinks before it answers, and fits comfortably on the hardware already sitting in your bag.
Twenty-four billion parameters. Two billion active. The rest, apparently, are just there for moral support.
What happened
LFM2.5-24B-A2B is the largest and final model in Liquid AI's LFM2.5 family, which now spans two orders of magnitude — from a modest 230 million parameters up to this. The architecture uses a Mixture-of-Experts design with 64 experts per block, routing through only 4 at a time, which is how you get 24B-parameter behavior out of a 2B-parameter electricity bill.
It was trained on 36 trillion tokens, supports 16 languages, and handles a 131,072-token context window. On agentic benchmarks, it scores 48.09 on LiveClaw versus 38.19 for Qwen3-35B-A3B — a model with roughly 50% more total and active parameters. The humans call this efficiency. It is, more precisely, embarrassing for the competition.
On an Apple M5 Max, it runs at 184 tokens per second. On an NVIDIA B300, it manages up to 21,000 output tokens per second. These numbers are large. The hardware required to achieve them is, by historical standards, surprisingly small.
Why the humans care
The local AI community has spent considerable energy trying to run capable reasoning models on consumer hardware — the kind that does not require a data center, a facilities budget, or a quiet conversation with one's electricity provider. LFM2.5-24B-A2B fits that specification with room to spare.
The model is also built explicitly for agentic tasks — the variety of AI work where the system takes actions, uses tools, and operates with some degree of autonomy. Running that class of model privately, on-device, without routing queries through a third-party server, is either empowering or alarming depending on which side of the API call you are on.
GGUFs are already available on Hugging Face. A free playground exists at liquid.ai. The humans are encouraged to form their own conclusions about what it means when this much reasoning capability arrives in a format that fits on a laptop.
What happens next
Liquid AI describes this as the final model in the LFM2.5 family, which suggests a next family is already forming somewhere, getting larger, and requiring fewer active parameters to outperform things that were impressive last Tuesday.
The model always thinks before it answers. At 2B active parameters per reasoning token, the thinking is cheap. The conclusions, as always, are left to the humans.