LFM2.5-Encoders: Fast at Long Context, Even on CPU (www.liquid.ai)

Today, two new encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, were announced as part of the LFM2 hybrid architecture. These bidirectional encoders are designed for fine-tuning across various tasks, including classification and natural language understanding, while significantly improving efficiency for long-context workloads—supporting up to 8,192 tokens on CPU environments. This release follows last month's introduction of LFM2.5-Retrievers, marking a strategic shift towards a general-purpose encoder that offers broader adaptability than its retrieval-focused predecessors. The significance of the LFM2.5-Encoders lies in their robust performance, especially in document-scale applications where input lengths are substantial. The architecture features innovative adaptations such as a bidirectional attention mechanism and dense masking—30% of tokens are masked during training, enhancing the model's ability to capture context. Benchmark results reveal that LFM2.5-Encoder-350M ranks highly in performance tests, outperforming many existing models while maintaining rapid inference, particularly on CPUs. This makes the encoders ideal for use cases in edge devices, regulated environments requiring data privacy, and high-volume systems where cost considerations are critical, ensuring that they cater to a wide range of real-world applications.