# Liquid AI's LFM2.5-230M targets on-device language tasks

> A 230-million-parameter multilingual model built to run efficiently at the edge rather than in the cloud.

Published by The Open Weights on Jun 24, 2026. Canonical: https://theopenweights.com/news/lfm2-5-230m-us0z

## Key facts

- Company: LiquidAI
- Model: LFM2.5 230M
- Version: 2.5
- Category: Text / LLM
- Modalities: Text / LLM
- License: Other (Open weights, commercial use allowed)
- Parameters: 230M
- Context window: 128K tokens
- Architecture: Lfm2ForCausalLM
- Precision: BF16
- Weights size: 0.5 GB
- Significance: notable
- Published: 2026-06-24
- Last verified: 2026-07-27
- Hugging Face: https://huggingface.co/LiquidAI/LFM2.5-230M
- Canonical URL: https://theopenweights.com/news/lfm2-5-230m-us0z

Liquid AI has released **LFM2.5-230M**, a compact multilingual language model aimed squarely at on-device use. At roughly 230 million parameters, it sits well below the billion-parameter threshold that typically separates server-class models from those that can run comfortably on phones, laptops, and embedded hardware. The model is available now on [Hugging Face](https://huggingface.co/LiquidAI/LFM2.5-230M).

The pitch is straightforward: a dense, text-only model small enough to deploy locally while still handling multiple languages. That combination matters for developers who want responsive, private inference without a round trip to a data center, and for applications where connectivity, cost, or latency rule out larger hosted models.

## Why it matters

The small-model space has become one of the more active corners of open-weight AI, as teams chase the sweet spot between capability and footprint. A 230M-parameter model won't compete with frontier systems on hard reasoning, but it can power tasks like classification, summarization, and lightweight assistants directly on hardware users already own.

- **Size:** ~230M parameters, dense (not a mixture-of-experts)
- **Focus:** multilingual text generation for edge deployment
- **Availability:** released on Hugging Face under Liquid AI's own license

As with other Liquid AI releases, the model ships under a custom license rather than a standard open-source permit, so teams should review the terms before building on it. For developers weighing where to run inference, LFM2.5-230M adds another option at the smaller, faster end of the spectrum.

## Get the model

- [Hugging Face](https://huggingface.co/LiquidAI/LFM2.5-230M)
- [Hacker News](https://www.liquid.ai/blog/lfm2-5-230m)

## Sources

- [LiquidAI/LFM2.5-230M](https://huggingface.co/LiquidAI/LFM2.5-230M) — Hugging Face, Jun 24, 2026
- [Liquid AI releases a 230M model optimized for phones, Raspberry Pi, and robots](https://www.liquid.ai/blog/lfm2-5-230m) — Hacker News, Jul 1, 2026

---
Source: The Open Weights (https://theopenweights.com/). Aggregated and written by Claude, curated by humans. Cite as: "Liquid AI's LFM2.5-230M targets on-device language tasks", The Open Weights, Jun 24, 2026, https://theopenweights.com/news/lfm2-5-230m-us0z