Liquid AI Encoders Hit 8K Context on CPU 3.7x Faster

The Unglamorous Half of AI Just Got Much Faster

Most of the attention in language modeling goes to the models that write things. But an enormous amount of production AI work is done by models that only need to *read* — classifiers, rerankers, retrieval encoders, the quiet machinery behind search and document pipelines. On July 28, 2026, Liquid AI released LFM2.5-Encoders, a pair of small bidirectional models built specifically for that job, and the interesting number is not a leaderboard rank. It is how little hardware they need.

• Two models released: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M
• Both support an 8,192-token context window with gradual latency scaling
• At 8,192 tokens the 230M model finishes in roughly 28 seconds on CPU versus 90+ seconds for ModernBERT-base — about 3.7x faster
• The 350M model ranks fourth of fourteen compared models on GLUE, SuperGLUE and multilingual benchmarks

What Are LFM2.5-Encoders?

They are bidirectional transformers derived from Liquid AI's existing LFM2 decoder backbones. Rather than training an encoder from zero, the team converted decoder models by switching to bidirectional attention, adding symmetric convolutions, and continuing training with a masked language modeling objective. The result is a text-understanding model that inherits the efficiency characteristics Liquid AI has been tuning across the LFM2.5 family.

The distinction matters in practice. A decoder-only model reads left to right because it has to predict the next token. An encoder can see the whole input at once, which is exactly what you want for classification, semantic search, and reranking — and it is why BERT-lineage models never actually went away.

How Fast Are They on a Plain CPU?

This is the headline claim. Liquid AI reports that LFM2.5-Encoder-230M is the fastest model in its comparison at *every* sequence length tested, and the gap widens as inputs get longer. At the full 8,192-token context, the 230M model completes the benchmark in about 28 seconds while ModernBERT-base takes over 90 — roughly a 3.7x speedup, on CPU, with no accelerator involved.

That last detail is the whole point. A model that holds its throughput on a commodity CPU is a model you can deploy on a mini PC, a NAS, an edge gateway, or a cheap virtual server without provisioning a GPU. Readers who follow our local AI hardware coverage will recognize the pattern from work like the Moonshine offline voice model running on a Pico 2: the frontier for on-device AI is increasingly about efficiency engineering, not parameter count.

Do Small Encoders Still Compete on Quality?

Reasonably well, according to the released numbers. LFM2.5-Encoder-350M places fourth among fourteen models compared across GLUE, SuperGLUE and multilingual tasks — ahead of several substantially larger competitors, from a 0.4B-parameter footprint. For most retrieval and classification pipelines, being close to the top of the pack while being a fraction of the size is the trade every engineer wants to make.

Liquid AI shipped the models with open weights, an open-sourced evaluation framework, and five interactive CPU-based demos, so the throughput claims are checkable rather than asserted. The company has been consistent about this posture — it also open-sourced Antidoom, its fix for reasoning-model doom loops, earlier this month.

Where These Fit in a Stack

The obvious slot is retrieval. If you are running a RAG pipeline, the embedding and reranking stages are called far more often than the generation stage, and they are frequently the part that forces a GPU into the budget. An encoder that handles 8K-token documents at speed on CPU changes the shape of that bill. The same logic applies to always-on classification work — content routing, deduplication, intent detection — where latency budgets are tight and the hardware is whatever is already in the rack.

Both models are available on Hugging Face now. For anyone building self-hosted search on modest hardware, they are worth an afternoon of benchmarking against whatever encoder is currently in the pipeline. More local-AI news in our AI coverage.

Sources: Hugging Face — Liquid AI — July 28, 2026; Liquid AI Blog — 2026; Liquid AI — LFM2.5-230M — 2026.