Liquid AI Releases LFM2.5-Encoder-230M and LFM2.5-Encoder-350M: Bidirectional Encoders That Stay Fast at 8K Context on CPU
Liquid AI ships 230M and 350M bidirectional encoders that handle 8K context on CPU — a rare win for efficiency without sacrificing benchmark rank.
Why it matters
Open-weight encoder models that combine long context (8K tokens) with CPU-feasible latency (28 seconds per pass on the smaller variant) challenge the assumption that capability requires scale or specialized hardware. Practitioners building search, semantic similarity, or embedding pipelines on resource-constrained infrastructure now have a credible alternative to larger or GPU-dependent encoders.
The key facts
12 to knowLFM2.5-Encoder-230M and LFM2.5-Encoder-350M released as open-weight
8,192-token context window on both models
LFM2.5-Encoder-350M ranks 4th of 14 models on GLUE/SuperGLUE/multilingual eval suite
LFM2.5-Encoder-230M completes 8K-token forward pass in ~28 seconds on CPU
Built on LFM2 hybrid backbone
Bidirectional encoder architecture (not autoregressive)
Liquid AI released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M as open-weight models
Both models support 8,192-token context window
350M ranks 4th of 14 on combined GLUE, SuperGLUE, and multilingual benchmarks
230M completes one 8K-token forward pass on CPU in ~28 seconds
Models built on LFM2 hybrid backbone
Open-weight release implies availability for practitioners to run locally or on-prem
Go to the source
MarkTechPostmarktechpost.com
Publisher excerpt: Liquid AI released two open-weight bidirectional encoders, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. Both carry an 8,192-token context and are built on the LFM2 hybrid backbone. The 350M ranks fourth of 14 models on a 17-task GLUE, SuperGLUE, and multilingual suite, behind only larger models.…