Liquid AI Releases LFM2.5-Encoder Models Claiming 3.7x Faster CPU Inference Than ModernBERT at Long Context

Liquid AI released two open-weight encoder models on Hugging Face on July 28, 2026: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. Both target high-volume text understanding tasks — classification, routing, PII detection, and policy enforcement — designed to run cheaply on CPU at long context lengths.

The headline claim: LFM2.5-Encoder-230M processes an 8,192-token input in about 28 seconds on CPU, versus over 90 seconds for ModernBERT-base, according to Liquid AI. That's roughly a 3.7x speedup at maximum context length. The company says its encoders show the largest advantage on CPU, where ModernBERT's throughput drops sharply as inputs grow while the LFM2.5-Encoders' latency degrades much more gradually.

Architecture and training

Both models are derived from Liquid AI's LFM2.5 decoder backbones (LFM2.5-230M and LFM2.5-350M), converted from causal to bidirectional encoders through three changes: a bidirectional attention mask so tokens attend to both left and right context, non-causal short convolutions with symmetric padding, and a masked-language-modeling objective that masks 30% of tokens during training.

Training ran in two stages. The first stage trained general language competence using masked-language modeling on a large web corpus at 1,024-token context. The second stage extended context to 8,192 tokens across the full data mix, with additional emphasis on factual, legal, and multilingual text.

Benchmark results

Liquid AI evaluated 14 models across 17 tasks pulled from GLUE, SuperGLUE, and multilingual classification benchmarks, fully fine-tuning each model per task and averaging scores across five held-out seeds. According to the company's published results, LFM2.5-Encoder-350M ranked fourth of the 14 models tested — behind three larger models, including one with roughly 3.5B parameters, nearly 10 times its size. LFM2.5-Encoder-230M reportedly outperformed ModernBERT-base and every EuroBERT variant tested, despite being smaller than most of them. Both new encoders scored above Liquid AI's own LFM2.5-Retrievers model, released the prior month, on these same tasks. The full benchmark framework and raw results have been open-sourced by the company.

On GPU, Liquid AI reports a smaller but similar pattern: ModernBERT-base leads on inputs below roughly 1,000 tokens, while the LFM2.5-Encoders take over from about 2,000 tokens onward.

Availability

Both models are open-weight and available now on Hugging Face under the LiquidAI organization, compatible with the transformers library (AutoModelForMaskedLM and AutoModel). Liquid AI published five accompanying demo Spaces covering zero-shot prompt routing, zero-shot policy linting, spell checking, PII detection across 40 categories and 16 languages, and a masked-diffusion text generation demo. The company also published a fine-tuning tutorial covering adaptation to long legal documents at 8k context. Pricing is not applicable, as these are open-weight downloadable models rather than hosted API products.

What this means

Encoder models like BERT and ModernBERT remain the workhorses behind production classification and routing systems precisely because they're cheap to run at scale, often on CPU rather than GPU. Liquid AI's pitch here isn't raw accuracy — the 350M model still trails three larger competitors on the benchmark suite — but cost efficiency at long context, a regime where older encoder architectures degrade badly on CPU. If the claimed 3.7x CPU speedup at 8,192 tokens holds up in independent testing, it directly lowers the compute bill for teams running always-on document classifiers, PII scanners, or intent routers on existing hardware rather than provisioning GPUs. The real test will be third-party reproduction of these benchmark numbers and latency figures outside Liquid AI's own framework.