Liquid AI releases LFM2.5 Encoder 230M and 350M

Liquid AI has launched two new encoder models, LFM2.5-Encoder-230M and LFM2.5-Encoder-350M, and published them on Hugging Face. The models are presented as bidirectional masked language models for fill-mask and broader text understanding workloads, with the main pitch being strong multilingual capability, long-context handling, and fast CPU-side performance. For developers building retrieval or classification systems outside large GPU setups, that combination is the main point of interest.

Confirmed

• Liquid AI released two models in the series: LFM2.5-Encoder-230M and LFM2.5-Encoder-350M.
• Hugging Face listings describe them as fill-mask models and bidirectional encoders / masked language models.
• Posts by @helloiamleonie and @tomaarsen say the models are adapted from the LFM2.5 backbone, with the key architectural change being a switch from causal attention to bidirectional attention.
• According to @pmttyji, the models target classification, sequence labeling, retrieval, reranking, semantic similarity, NLI, and extractive QA, and support 15 languages.
• According to @maximelabonne, the 230M version supports an 8,192 context length.
• Multiple posts relay Liquid AI's positioning that the models train quickly and are optimized for long-context speed on CPU.

Why it matters

• The release expands the LFM2.5 family into encoder-style workloads rather than only generative use cases.
• At 230M and 350M parameters, the models are relatively lightweight, while still being pitched for multilingual retrieval and understanding over long inputs.
• If the claimed CPU efficiency holds in practice, the models could be useful for local or cost-sensitive deployments where reranking, search, and document understanding matter more than text generation.

2026-07-27 ~ 2026-07-29 · 7 related posts

• [source] LiquidAI posts LFM2.5-Encoder-350M on Hugging Face as a new fill-mask encoder — LiquidAI · 2026-07-27
• 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
• Liquid AI’s LFM2.5 encoders bring 8k multilingual retrieval to 230M and 350M models — pmttyji · 2026-07-29
• [source] LiquidAI’s 230M LFM2.5 encoder trends on Hugging Face — LiquidAI · 2026-07-29