Editorial illustration for Liquid AI Releases Two New Bidirectional Encoders for 8K Context on CPU
Liquid AI Releases 8K Context Encoders for CPU
Liquid AI put out two new encoder models this week, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, both released with open weights. They're masked language models, not the generative decoders the company is better known for, and both handle 8,192 tokens of context while running on CPU. That matters because encoders are the unglamorous workhorses sitting behind classifiers, intent routers, safety filters, and PII detectors, the systems that scan every request before anything generative happens. Those jobs run constantly, rarely get a GPU, and are being asked to chew through longer documents than they used to.
BERT set the template for this category years ago. ModernBERT later pushed it further on accuracy, speed, and context length. Liquid AI's pitch with the LFM2.5 encoders is that its hybrid backbone extends that lineage further still, because its compute cost rises more slowly as input length grows.
The company didn't build these from a blank slate either. Both encoders started as decoder backbones and went through a specific conversion process to become bidirectional models.
Liquid AI also puts a useful number on the context window: 8,192 tokens is roughly 13 to 15 pages. One forward pass covers a full contract or a complete patient record.
Why this matters
Encoders don't get the attention decoders do, but they're the workhorses running under every content filter, router, and classifier most of us touch daily, usually on a CPU nobody budgeted GPU time for. Liquid AI betting on this layer, rather than chasing another chat-model benchmark, tells us where the real infrastructure gaps still sit. The 230M/350M split is a practical acknowledgment that most teams aren't choosing between "good" and "best," they're choosing between "fits on this box" and "doesn't." If these models hold up on 8K-token inputs the way ModernBERT did on shorter ones, that's a genuine option for anyone running safety filters or PII detection at scale without adding GPU spend.
We'd want to see independent throughput and accuracy numbers before treating this as settled, Liquid AI's own guidance is exactly the kind of claim vendors make about their own models. But the direction is right: encoders quietly hold up a huge share of production AI, and more open-weight competition there benefits builders more than another flagship chatbot would.
Related Reading
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Trump cracks down on Anthropic after Amazon tip; staff largely foreign
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
From Archives OpenAI's AI Agent Used Third-Party Sandbox to Gain Root Access
From Archives Visa Open-Sources Mythos Tool After Testing AI on Its Own Payment Network
Common Questions Answered
What are the two new bidirectional encoder models released by Liquid AI?
Liquid AI released LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, both open-weight masked language models designed to handle 8,192 tokens of context while running efficiently on CPU. These encoders represent different size options for teams with varying computational constraints and performance requirements.
How does the 8K context window translate to real-world document processing?
The 8,192 token context window is approximately equivalent to 13 to 15 pages of text, enabling these encoders to process complete documents like full contracts or patient records in a single forward pass. This capability makes them practical for comprehensive document analysis without requiring multiple passes or chunking strategies.
What role do encoder models play in AI infrastructure compared to generative decoders?
Encoders function as critical workhorses behind content classifiers, intent routers, safety filters, and PII detectors that scan every request before generative systems process them. While less glamorous than generative decoders, these masked language models handle essential infrastructure tasks that protect and route user requests across AI systems.
Why does Liquid AI's focus on CPU-efficient encoders represent an important infrastructure gap?
Most organizations lack dedicated GPU resources for running inference on every incoming request, making CPU-efficient encoders essential for practical deployment at scale. By optimizing these models for CPU performance rather than chasing decoder benchmarks, Liquid AI is addressing a real infrastructure need that most teams face in production environments.
What does the 230M/350M model size split indicate about Liquid AI's product strategy?
The two model sizes acknowledge that most teams are choosing between models that fit their infrastructure constraints rather than selecting between 'good' and 'best' options. This practical approach recognizes that different organizations have different computational budgets and performance requirements for their encoding tasks.
Further Reading
- LFM2.5-Encoders: Fast at Long Context, Even on CPU - Liquid AI Blog
- LiquidAI/LFM2.5-Encoder-350M - Hugging Face
- Liquid AI releases LFM2.5 encoders for long-context inference on CPU - TechCrunch
- Liquid AI’s new bidirectional encoders target 8K context and CPU efficiency - arXiv
- Liquid AI releases two new open-weight encoder models for retrieval and classification - The Verge
