Liquid AI Skips Benchmark Hype for 24B On-Device Model

Liquid AI released LFM2.5-24B-A2B Tuesday, a 24-billion-parameter on-device model designed to run private agents without sending data to the cloud.

Key Takeaways

Liquid AI released LFM2.5-24B-A2B Tuesday, a 24-billion-parameter on-device model designed for private agents without cloud data transfers
The openly downloadable Hugging Face model uses mixture-of-experts architecture, activating only a fraction of its parameters for each task
On-device deployment keeps users’ prompts, files and screen contents from going to an external server
Liquid AI has not published independent third-party benchmark comparisons alongside the release

The company detailed the release in a Hugging Face blog post Tuesday, positioning it for developers building agents on laptops, phones or edge servers rather than through a cloud API.

The “A2B” name signals a mixture-of-experts design, the model contains specialized sub-networks but activates only a small subset for a given task, reducing the computing load compared with running all 24 billion parameters every time.

Why Running A Model Locally Changes What It Can Do

For LFM2.5-24B-A2B, on-device deployment means a user’s prompts, files or screen contents do not go to an external server, which matters for anyone handling sensitive documents, health data or proprietary business information through an AI agent.

Cloud-hosted models require trusting a third party’s data handling and accepting network latency on every request.

The release is an openly downloadable Hugging Face model rather than a gated API product, so it does not establish cloud API pricing, rate limits, regions or tiers.

Its mixture-of-experts architecture is the mechanism Liquid AI is using to target local deployment, only a fraction of the model’s parameters activate at once, rather than the full 24 billion.

Also Read: Alibaba’s Qwen Slashes Image Model To 8 Steps, Skips Quality Benchmarks

Liquid AI’s Shift From Research Lab To Deployment-Focused Model Maker

The company spun out of MIT research into liquid neural networks, an architecture distinct from the transformer designs used by most large labs.

Its LFM (Liquid Foundation Model) series has targeted efficiency and on-device performance rather than leaderboard-topping benchmark scores, a different strategy from labs that compete primarily on raw capability. Tuesday’s LFM2.5-24B-A2B release continues that approach as an openly downloadable model on Hugging Face.

The Open-Weight Race For The Edge

The launch demonstrates that Liquid AI is shipping a 24-billion-parameter model for local use, but a download alone does not show how it performs during ordinary agent use with real files, screens and multi-step tasks.

The test for LFM2.5-24B-A2B will be adoption numbers on Hugging Face over the coming weeks and whether developers report it holding up against cloud-hosted alternatives on real agentic tasks rather than synthetic benchmarks.

The company has not published independent third-party benchmark comparisons alongside the release.