Liquid AI Launches LFM2.5-24B-A2B to Run AI Agents Directly on Computers

Liquid AI launches LFM2.5-24B-A2B, a 24-billion-parameter AI model with 2 billion active parameters, capable of running agents locally at high speed.

Running an AI agent directly on one's computer, without sending data to a remote server, all while maintaining strong performance on complex tasks. This is what LFM2.5-24B-A2B, Liquid AI's new reasoning model, offers. With 24 billion total parameters, but only 2 billion activated for each processed token, it prioritizes execution speed and resource efficiency. A 24-billion-parameter model, with 2 billion active

LFM2.5-24B-A2B is the largest model in the LFM2.5 family, which now ranges from 230 million to 24 billion parameters.

Its Mixture-of-Experts (MoE) architecture mobilizes a fraction of its capabilities at each step rather than the entire model. This design reduces compute requirements while maintaining performance suitable for agents capable of reasoning, using tools, and chaining multiple operations.

Compared to LFM2-24B-A2B, its predecessor, the context window expands from 32,768 to 131,072 tokens, while the internal vocabulary doubles to 128,000 tokens, notably improving the processing of languages that use non-Latin writing systems.

Liquid AI also increased the pre-training volume from 17 to 33 trillion tokens, before continuing its training on reasoning, coding, and tool-use tasks. Unlike the previous generation, the model systematically produces an explicit reasoning phase before its response. Competitive performance against larger models

Liquid AI compared its model to Qwen3.6-35B-A3B, Gemma 4 26B, and Nemotron 3.5 Lightning across 16 benchmarks.

On the eight tests dedicated to agents and tool use, LFM2.5-24B-A2B comes out on top or tied on five of them. Notably, it scores 48.09 on LiveClaw, compared to 38.19 for Qwen3.6-35B-A3B, and 27.69 on AutomationBench, compared to 8.72 for the same competitor.

The model also takes the lead on the three instruction-following benchmarks. However, results are less favorable on certain knowledge, coding, and web navigation tasks, where larger models retain an edge.

These comparisons come from benchmarks published by Liquid AI and do not guarantee identical performance in all environments. Up to 184 tokens per second on a Mac

Speed is the model's other selling point. On a computer equipped with an Apple M5 Max chip, Liquid AI measures 184 generated tokens per second, compared to 114 on an AMD Ryzen AI Max+ 395. In the compared configurations, the model is reported to be up to 1.7 times faster than its MoE competitors while using less memory.

On an NVIDIA B300 card, throughput exceeds 21,000 output tokens per second when processing many requests concurrently. This refers to cumulative server throughput, distinct from the speed achieved by a user on their computer.

Liquid AI also illustrates local execution with Hermes Agent, installed on an M5 Max MacBook Pro. The agent can plan a task, call tools, and produce a synthesis without relying on a remote generation API. Open weights and integration with existing tools

The pre-trained and fine-tuned versions of LFM2.5-24B-A2B are available on Hugging Face, with the ability to download, customize, and run them on one's own infrastructure, subject to the model's license terms.

Support for llama.cpp, MLX, vLLM, SGLang, and ONNX facilitates its use across different hardware setups. An API is also offered via Liquid Console, at $0.10 per million input tokens and $1.00 per million output tokens.

With this latest iteration of LFM2.5, Liquid AI continues an approach where efficiency matters as much as model size. For developers, the appeal lies particularly in the ability to run agents directly on their machines, with greater control over data and without systematically relying on a cloud service. Capabilities remain dependent on the hardware used, but the gap with hosted solutions continues to narrow on certain specialized tasks.