Liquid AI's LFM2.5-24B-A2B: an open-weight 24B reasoning model built for on-device agents

Liquid AI announced on October 13, 2026, the release of LFM2.5-24B-A2B, the largest model in its LFM2.5 family and a clear sign of where the on-device AI race is heading. Rather than chasing raw benchmark bragging rights, it is a reasoning model deliberately designed to fit on the hardware you already have. That makes it a genuinely interesting release for anyone watching the open-weight space.

A 24B model built to run on your laptop

The model carries 24.0B total parameters but activates only 2.45B of them at a time, a mixture-of-experts design with 40 layers, 64 experts per block, and top-4 routing. In practical terms, it is a large knowledge-bearing model that costs like a much smaller one to run. It supports a 131,072-token context, 16 languages, and ships as a reasoning-only model that emits an explicit chain of thought before each answer. Checkpoints are published as open weights on Hugging Face, with FP8, GGUF, ONNX, and MLX-4bit formats and day-one support for llama.cpp, MLX, vLLM, and SGLang, so the whole point is local, hands-on inference.

What this means for you

If you build local agents, LFM2.5-24B-A2B is a strong option: Liquid AI reports 184 tokens per second on an Apple M5 Max and 114 on an AMD Ryzen AI Max+ with Q4 quantization, at just 15 GB on the M5 Max, which it claims is the least memory of the compared group. Its benchmarks lean hard into tool use and agentic workflows, where it claims a standout 27.69 on AutomationBench, more than three times the next best. The honest tradeoff is knowledge and coding, the larger comparison models still edge it on GPQA Diamond, HLE, and SWE-Bench, and Liquid AI itself recommends pairing it with retrieval or web search for knowledge-heavy work.

• It runs on-device: 24B total but only about 2B active parameters, with llama.cpp and MLX support and roughly 15 GB of memory on high-end silicon
• It is open-weight and local-first: checkpoints in native, FP8, GGUF, ONNX, and MLX-4bit, plus a 131,072-token context and 16-language support
• It is tuned for agent tool use rather than raw knowledge: strong reported scores on tool-use and instruction-following benchmarks, with a chain-of-thought reasoning mode
• The tradeoff is knowledge-heavy and some coding benchmarks, where larger models still lead, and it is best paired with retrieval or web search

Sources: Unite AI - Liquid AI Releases Open-Weight LFM2.5-24B-A2B for On-Device Agents; Hugging Face - LFM2.5-24B-A2B model card; Liquid AI - LFM2.5-24B-A2B release blog.

Corrections: Report an error at corrections@wireshaman.com.