Liquid AI releases LFM2.5 24B A2B, an open weight model with 2 billion active parameters built to run agents on laptops and PCs
Today we release 饾櫥饾櫟饾櫦饾煾.饾熁-饾煾饾熀饾櫛-饾櫚饾煾饾櫛, the largest 饾櫥饾櫟饾櫦饾煾.饾熁 family member. It is an efficient local model for laptops and PCs to work through multi-step tasks inside popular agent harnesses. > 24B total parameters, 2B active per token > Context expanded to 131K tokens > A 128K-token vocabulary > 33T pretraining tokens > It is a reasoning model > Very easy to fine-tune and customize > Performs competitively with Qwen3.6-35B-A3B, a model with ~50% more total and active parameters, on agentic and tool-use tasks. > Open base and post-trained weights on @huggingface 馃 1/
Context
Small, efficient open models that run agents locally could cut costs and keep data on users' devices, a growing alternative to cloud based frontier models.
Liquid AI is an MIT spinout that builds efficient models, and this is its own release with a blog post and Hugging Face weights. Benchmark comparisons in the post are the company's own. The model's full name is LFM2.5 24B A2B.
Deep dive
On October 13, Liquid AI posted that it was releasing the largest member of its LFM2.5 family, an efficient local model for laptops and PCs that works through multi step tasks inside popular agent harnesses.
The company listed 24 billion total parameters with 2 billion active per token, a context window expanded to 131,000 tokens, a 128,000 token vocabulary and 33 trillion pretraining tokens. It said the model is a reasoning model and very easy to fine tune and customize.
Liquid AI wrote that the model performs competitively with Qwen3.6 35B A3B, which has about 50% more total and active parameters, on agentic and tool use tasks. The attached chart compares it with Qwen, Gemma and Nemotron models on 16 benchmarks, where it leads on some and trails on others.
The company said both base and post trained weights are open on Hugging Face. According to Liquid's blog and documentation, builds are available in several formats, including GGUF and MLX, and the model works with common local tools such as llama.cpp, MLX, vLLM and SGLang.
Because only 2 billion parameters are active per token, the model can run quickly on consumer hardware, which is the core of Liquid's pitch.
What to expect next
Watch independent benchmarks and developer reports on how well the model handles real agent tasks locally compared with Qwen and Gemma models of similar size.
Also watch whether laptop and PC makers ship it on devices, and what license terms apply to commercial use of the open weights. Downloads and community fine tunes on Hugging Face in the coming weeks will be an early sign of whether developers adopt it for local agents.