Liquid AI Releases LFM2.5 Encoders for Fast CPU Long-Context Processing

Liquid AI has released two brand-new bidirectional encoder models: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M. Modified from the LFM2.5 architecture, these models primarily target fill-mask tasks, emphasizing outstanding performance in multilingual and long-context scenarios, alongside highly efficient CPU inference speeds. Both models are now open-sourced on Hugging Face.

已确认

• 模型规格: Released in two parameter sizes: 230M and 350M. According to @maximelabonne, the 230M version supports a context length of 8,192.
• 架构调整: @helloiamleonie and @tomaarsen noted that the new models are based on the LFM2.5 backbone, with the key change being a shift from the previous causal attention mechanism to a bidirectional attention mechanism.
• 核心优势: The official team and multiple authors emphasized that these two models train quickly, possess strong multilingual capabilities, and are specifically optimized for long-context processing speeds on CPUs.
• 发布渠道: Both models are available on Hugging Face, with the 350M version's model card explicitly labeling it as a bidirectional masked language model.

为什么重要

• This release fills the gap in the LFM2.5 series for pure encoders. By focusing on CPU-side long-text processing speeds, these lightweight models provide developers with a new option for handling multilingual retrieval and comprehension tasks in resource-constrained local environments.

2026-07-27 ~ 2026-07-29 · 5 related posts

• [source] LiquidAI posts LFM2.5-Encoder-350M on Hugging Face as a new fill-mask encoder — LiquidAI · 2026-07-27
• [source] LiquidAI launches two multilingual encoder models with CPU speed and long-context gains — maximelabonne · 2026-07-28
• Liquid AI releases LFM2.5 encoders for long-context tasks on CPU — tomaarsen · 2026-07-29
• LiquidAI releases LFM2.5-Encoder-350M on Hugging Face — tomaarsen · 2026-07-29
• [source] Liquid AI launches 230M and 350M encoders that stay fast on long context — helloiamleonie · 2026-07-29