LiquidAI releases LFM2.5-Encoder in 250M and 350M sizes

On October 3, 2026, a post on the LocalLLaMA subreddit reported that LiquidAI has released the LFM2.5-Encoder series, which consists of two bidirectional encoders, at 250M and 350M. That post, published at 15:00 that day, is the only account of the release available so far.

According to the post, the series is built on the LFM2 architecture. The models are masked language models with full bidirectional attention, cover 15 languages, and are designed to run efficiently on-device. The post quotes the official description of LFM2.5-Encoder-350M, which it calls a multilingual bidirectional encoder and the larger of the two, positioned for maximum downstream quality. Like the rest of the series, it is a masked language model with full bidirectional attention built on LFM2, intended to be fine-tuned into task-specific models across classification, token classification, retrieval, reranking, and semantic similarity, and to run efficiently on-device.

On performance, the post says the company describes the models as comparable to the best encoders of the same size. No benchmarks, comparison targets, or figures were provided, and the post does not spell out how the 250M and 350M versions differ in detail. It also does not say whether model weights, a license, or a download link accompany the release.

As it stands, the story is at the announcement and description stage. No follow-up developments have been reported.