LFM2.5-24B-A2B: An On-Device Agent Model With 2B Active Params

On AutomationBench, Liquid AI’s new LFM2.5-24B-A2B scored 27.69, more than three times the 8.72 of Qwen3.6-35B-A3B, a model with about 50% more total and active parameters. Liquid AI’s launch post pitches it as a reasoning model for “fast, private agents on-device”, with only 2B of its 24B parameters active per token. For developers who want agents that call tools on a laptop instead of a cloud API, it’s a strong option, with a clear weak spot in coding.

What Happened

LFM2.5-24B-A2B keeps the architecture of its predecessor: 40 layers, 64 experts per mixture-of-experts block with top-4 routing, and a hybrid of gated short convolutions and grouped-query attention. The context window grows from 32K to 131K tokens and the vocabulary doubles to 128K tokens. Unlike the earlier model, it always produces a chain of thought before answering. Pre-training was scaled from about 17T to 33T tokens, and post-training ended with reinforcement learning inside agent harnesses including OpenClaw, Hermes Agent, Pi, and OpenCode.

The model card and launch post list day-one support for llama.cpp, MLX, vLLM, SGLang, and ONNX. Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 on a Ryzen AI Max+ 395, up to 1.7x faster than the next-best compared model while using the least memory, and over 21K output tokens per second at high concurrency on a single NVIDIA B300.

Why It Matters

The benchmarks split cleanly. Against Qwen3.6-35B-A3B, Gemma-4-26B-A4B-it, and Nemotron-3.5-Lightning, Liquid AI says the model leads or matches on five of eight tool-use and agentic benchmarks and on all three instruction-following tests:

Larger models keep an edge on knowledge-heavy tests and coding. Liquid AI’s own advice is to pair it with retrieval or web search for knowledge-heavy work. All figures are Liquid AI’s, run with vLLM using each benchmark’s recommended settings, and the post cites no independent reproduction.

What This Means for Builders

The economics are the draw. With only 2B active parameters, each reasoning token is cheap, which Liquid AI says makes it practical to run agents continuously on local hardware. Serving it behind any OpenAI-compatible endpoint lets existing harnesses such as Hermes Agent point at it with a few config changes, such as a custom base URL and a 131,072-token context length. It joins a growing set of local agent models, including Google’s Gemma 4 for on-device agentic AI, with tool-calling scores Liquid AI says rival larger models.

Source: Hugging Face Blog (Liquid AI)

Content created with AI assistance and reviewed for accuracy.