---
type: "post"
slug: "liquid-ai-releases-cpu-friendly-long-context-encoders"
title: "Liquid AI releases long-context encoders designed for CPU inference"
description: "The 230M and 350M LFM2.5 encoders target classification, routing, policy checks, retrieval, and multilingual PII detection."
retrieval_nugget: "The 230M and 350M LFM2.5 encoders target classification, routing, policy checks, retrieval, and multilingual PII detection."
published_at: "2026-08-23"
updated_at: "2026-08-23"
record_date: "2026-08-23"
date_kind: "scheduled_at"
topics: ["encoders","local-models","retrieval","privacy"]
entities: ["huggingface.co"]
editorial_format: "brief"
basket_id: "64af3bcb-1c2d-42a9-a664-91510a61d75a"
basket_revision: 1
source_urls: ["https://huggingface.co/blog/LiquidAI/lfm2-5-encoders"]
visual_decision: "text_only"
recovery_incident: "NR-2026-08-15-HERMES-SITE-COPY"
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:liquid-ai-releases-cpu-friendly-long-context-encoders"
status: "published"
visuals: []
editorial_provenance: {"schema_version":"newruntime-editorial-copy-v1","content_status":"source_grounded_final","final_copy_sha256":"sha256:86a0c3bf6992360af1010a536a3851c3097858030849a80cf6d7b32ef34982f7","reviewed_at":"2026-08-15T20:30:00.000Z","source_evidence_count":1,"verified_claim_count":2}
routes: {"html":"https://newruntime.com/posts/liquid-ai-releases-cpu-friendly-long-context-encoders/","markdown":"https://newruntime.com/posts/liquid-ai-releases-cpu-friendly-long-context-encoders.md","json":"https://newruntime.com/posts/liquid-ai-releases-cpu-friendly-long-context-encoders.json"}
---

# Liquid AI releases long-context encoders designed for CPU inference

## Retrieval answer

The 230M and 350M LFM2.5 encoders target classification, routing, policy checks, retrieval, and multilingual PII detection.

Liquid AI released LFM2.5-Encoder models with 230 million and 350 million parameters and context up to 8,192 tokens. The intended uses include classification, routing, policy checks, retrieval, and PII detection. The PII variant covers 40 entity types across 16 languages.

Liquid AI reports that its base encoder processes an 8K-token input on CPU in roughly 28 seconds, compared with more than 90 seconds for ModernBERT-base in the company's test. That is a vendor comparison, and downstream accuracy still depends on fine-tuning and the target data.

The product class is useful around agents because many checks do not require generation. A local encoder can route requests, detect sensitive text, or classify traces without sending data to a frontier model. Teams should measure recall carefully: a cheap privacy classifier is valuable only when missed entities and uncertain cases have an escalation path.

## Source

- [Liquid AI on Hugging Face](https://huggingface.co/blog/LiquidAI/lfm2-5-encoders)