Liquid AI has released LFM2.5-24B-A2B, an open-weight model with 24 billion parameters that the company says can hang with larger competitors. The twist is that it only uses a fraction of that size at any given moment.
The model went live on October 13, 2026, and is available on Hugging Face. It is the biggest entry in the LFM2.5 family. It also caps the series, which Liquid AI has spent the year aiming at one goal: making capable AI run locally on consumer hardware instead of in a distant data center.
Big model, small footprint
The headline spec is a split personality. LFM2.5-24B-A2B holds 24 billion parameters in total but activates only 2 billion per token.
That design is what makes the speed numbers possible. Liquid AI says the model reaches up to 184 tokens per second on Apple’s M5 Max chip. The company also claims the model runs up to 1.7 times faster than comparable mixture-of-experts models on edge devices.
The context window sits at 131,000 tokens. The vocabulary covers 128,000 entries, and the model supports 16 languages.
How it stacks up against bigger rivals
According to Liquid AI, LFM2.5-24B-A2B leads or ties on 5 out of 8 agentic and tool-use benchmarks when measured against larger models, including Qwen3.6-35B-A3B. That Qwen model carries 35 billion total parameters and 3 billion active ones.
The standout result is on AutomationBench, where the company reports a score of 27.69 versus 8.72.
Because the weights are open, independent verification can actually happen. Anyone can download the model from Hugging Face and run their own tests.
An upgrade over the original LFM2 version
This is not Liquid AI’s first 24-billion-parameter model. It builds on LFM2-24B-A2B, which shared the same overall size and active parameter count.
The context window jumped from 32K to 131K tokens. The vocabulary doubled to 128K. Liquid AI also put the model through extensive additional pretraining and reinforcement learning targeting reasoning and tool use specifically.
The LFM2.5 family, now complete
Liquid AI is an MIT spin-off, and the LFM2.5 series has been its main product push. The lineup spans models from 230 million parameters up to this new 24-billion-parameter release.
Earlier 2026 releases in the series focused on edge agent capabilities, meaning AI assistants that can act on a device without phoning home to the cloud. The 24B model sits at the top of that range as the most capable option, while the smaller versions cover lighter hardware.
Disclosure: This article was edited by Diego Almada Lopez. For more information on how we create and review content, see our Editorial Policy.