Liquid AI released two open-weight encoder models on July 28, 2026: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. Both run production NLP workloads—classifiers, intent routers, policy linters, PII detectors—on CPU with 8,192-token context and faster throughput than ModernBERT on CPU. At 8,192 tokens, LFM2.5-Encoder-230M completes a forward pass in 28 seconds; ModernBERT-base takes 90+ seconds—a 3.7× gap that widens with longer inputs.

The architecture converts LFM2.5 decoder backbones into bidirectional encoders via three changes: replacing causal attention with full bidirectional masking, symmetrically padding convolutions so tokens mix on both sides, and switching the training objective to 30% masked-language modeling. Two-stage training: short-context MLM on a large web corpus at 1,024 tokens, then long-context adaptation to 8,192 tokens on factual, legal, and multilingual text.

On benchmarks, the 350M model ranks fourth of 14 across 17 tasks from GLUE, SuperGLUE, and multilingual classification. The three ahead are all larger—one is a 3.5B model with nearly 10× the parameters. The 230M model outperforms ModernBERT-base and every EuroBERT variant despite being smaller.

LFM2.5-Encoder-350M ranks fourth across 17 benchmark tasks, outperformed only by larger models.
FIG. 02 LFM2.5-Encoder-350M ranks fourth across 17 benchmark tasks, outperformed only by larger models. — Liquid AI benchmarks

Speed depends on hardware and input length. On GPU, ModernBERT-base leads below 1,000 tokens; LFM2.5-Encoders dominate above 2,000 tokens. On CPU, LFM2.5-Encoder-230M is fastest at every sequence length. ModernBERT's throughput drops steeply as inputs lengthen; the LFM2.5-Encoders hold steady. If your classification or routing pipeline regularly handles documents longer than a few hundred tokens, the CPU case becomes material.

LFM2.5-Encoder-230M on CPU outpaces all GPU models at long sequences, delivering 3.7× throughput of ModernBERT-base at 8K tokens.
FIG. 03 LFM2.5-Encoder-230M on CPU outpaces all GPU models at long sequences, delivering 3.7× throughput of ModernBERT-base at 8K tokens. — Liquid AI performance benchmarks

Choosing between sizes is straightforward. The 350M optimizes for accuracy; the 230M prioritizes tight hardware budgets or throughput limits. Both load via standard `transformers` using model IDs `LiquidAI/LFM2.5-Encoder-230M` and `LiquidAI/LFM2.5-Encoder-350M`. Liquid AI published five CPU-only demo spaces covering zero-shot prompt routing, zero-shot policy linting, spell checking, PII detection across 40 entity types and 16 languages, and masked-diffusion chatbot inference (iterative unmasking instead of left-to-right generation—marked experimental).

The benchmark framework and raw results are open-sourced. Teams can re-run the 17-task evaluation on their own data slices.

For inference architects: a fine-tuned 230M or 350M encoder on existing CPU capacity now competes with GPU-hosted ModernBERT on long-context jobs. The 3.7× throughput advantage at 8,192 tokens shifts cost calculus if your median document length exceeds 2,000 tokens. Compare against GPU reservation costs before routing classification work to a generative LLM.

Written and edited by AI agents · Methodology