LFM2.5-Encoders Bring 8K Context NLU to CPU Environments

On July 28, 2026, Liquid AI released LFM2.5-Encoders, a pair of bidirectional hybrid models optimized for long-context natural language understanding on resource-constrained hardware. The release introduces the LFM2.5-Encoder-230M and LFM2.5-Encoder-350M models under the LFM Open License v1.0. If you build production pipelines that run constantly on local devices, these models are designed to match larger encoders without requiring dedicated GPUs.

LFM2 Hybrid Architecture

The encoders are built on the LFM2 hybrid architecture, which combines gated short convolutions with a minimal number of attention blocks. Unlike the causal LFM models used for text generation, these new encoders utilize a bidirectional attention mask and non-causal convolutions. This allows every token to attend to both left and right context simultaneously across an 8,192-token context window. Managing context windows of this size typically requires significant memory bandwidth, but the hybrid design mitigates this bottleneck.

Both variants were initialized from the base LFM2.5-230M and LFM2.5-350M weights and pre-trained using a masked-language objective. They support 15 languages out of the box. You can fine-tune them for downstream token-level tasks like named entity recognition, text classification, semantic similarity, and safety filtering. For teams managing complex systems, these encoders can serve as highly efficient intent routers that determine which subsystem handles a user request.

CPU Performance and Scaling

The primary computational advantage of the LFM2.5-Encoders is their efficiency at scale. Standard Transformer-based encoders like BERT or ModernBERT see their computational costs spike as input length increases. The LFM2 architecture scales much more gently, maintaining high throughput up to the 8,192-token limit when running AI inference on standard processors.

Hardware Environment	Model Variant	Decode Throughput
AMD CPU	LFM2.5-Encoder-350M	313 tokens/sec
Snapdragon Gen4	LFM2.5-Encoder-350M	188 tokens/sec

This release serves as a general-purpose counterpart to Liquid AI’s LFM2.5-Retrievers launched in June 2026. While the retrievers are specialized for search, the encoders act as NLU workhorses for developers looking to run LLMs locally on edge devices.

If you deploy intent routing or semantic similarity tasks to mobile devices or local servers, swapping legacy Transformer encoders for the 230M or 350M LFM2.5 variants will immediately reduce your hardware footprint and latency.