Liquid AI releases two lean encoder models that read long documents fast, even on a laptop
LFM2.5-Encoder-230M and LFM2.5-Encoder-350M can scan an entire contract in under 30 seconds on ordinary computer hardware, no specialist chip required.
Key points
- Liquid AI released two encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, on Hugging Face in 2026.
- The 230M model runs roughly 3.7 times faster than ModernBERT-base on a standard laptop CPU at long inputs.
- Both models handle up to 8,192 tokens, roughly six to eight pages of dense text, in a single pass.
- LFM2.5-Encoder-350M ranked fourth out of 14 models tested, beaten only by models up to ten times its size.
- Both models are open-weight and free to download from Hugging Face.
Liquid AI, a Boston-based AI company, has released two new encoder models, a type of AI model that reads and understands text rather than generating it. The models are called LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, available now on Hugging Face.
If you've never encountered an encoder, the distinction is simple. A chatbot like ChatGPT writes new sentences word by word. An encoder does the opposite: it reads text and decides what it means, whether that's labelling an email as spam, flagging personal data in a document, or routing a customer query to the right department. Businesses run these tasks constantly, which makes speed and cost the deciding factors.
Why does running on a laptop CPU matter?
Most capable AI models need expensive specialised chips called GPUs to run at useful speeds. These encoders don't. On an ordinary laptop processor, the 230M model processes a full 8,192-token document, roughly six to eight pages of text, in about 28 seconds. ModernBERT-base takes over 90 seconds on the same hardware for the same length.
That 3.7-times speed gap matters most for smaller teams that don't want to pay cloud GPU bills for routine classification work. We covered a similar push toward accessible hardware on 23 July 2026, when Hugging Face brought 4-bit image diffusion to consumer GPUs: the pattern of shrinking the hardware requirement without sacrificing performance is real and accelerating.
How good are the results?
Speed means little without accuracy. Liquid AI tested both models across 17 tasks from GLUE and SuperGLUE, two standard benchmark suites researchers use to score language understanding. Out of 14 models, LFM2.5-Encoder-350M placed fourth. Every model ahead of it was larger, including one nearly ten times its size at 3.5 billion parameters.
The 230M model beat ModernBERT-base and every EuroBERT variant, despite being smaller than most of them.
| Model | Parameters | GLUE/SuperGLUE rank |
|---|---|---|
| LFM2.5-Encoder-350M | 350M | 4th of 14 |
| LFM2.5-Encoder-230M | 230M | beats ModernBERT-base and all EuroBERT models |
| ModernBERT-base | unspecified | below both LFM2.5 models |
What can someone actually build with this?
Liquid AI's demos include a tool that spots and removes 40 types of personal information across 16 languages, a policy checker that scans text against company rules written in plain English, and zero-shot prompt routing that sends user requests to the right service automatically. Every demo runs in a browser with no setup, on CPU only.
Developers can download either model and adapt it to a specific task, a process called fine-tuning, using Hugging Face's standard tools. Liquid AI published a tutorial covering long legal documents.
Common questions
Do I need a powerful computer to use these models?
No. Both models run on a standard laptop or server CPU. A GPU isn't required, though using one will speed them up further at shorter documents.
Are these models free to use?
Yes. Both are open-weight, meaning anyone can download them at no cost, including for commercial projects, subject to the licence terms on Hugging Face.
Which size should I choose?
Pick the 350M model when accuracy is the priority. Choose the 230M model when throughput matters more or your hardware is limited.
The honest watch-point here is benchmarks. GLUE and SuperGLUE are well-understood, but they don't always predict how a model behaves on your specific documents. The speed numbers are compelling; teams should run their own data through both models before committing.
