Liquid AI ships a 350M encoder built for CPUs

Liquid AI has released LFM2.5-Encoder-350M, a compact bidirectional text encoder aimed at embedding workloads that need to run quickly and cheaply. At roughly 350 million parameters, it slots into the sub-1B tier where the priority is throughput and deployability rather than raw scale.

The model is a classic encoder — bidirectional attention that reads an entire sequence at once — rather than a generative decoder. That design makes it well suited to retrieval, classification, and semantic search, where the goal is to turn text into vectors rather than produce new tokens. Liquid AI frames the release around fast long-context inference on CPU, a nod to teams that want to avoid the cost and complexity of GPU serving.

Why it matters

Much of the open-weights conversation is dominated by ever-larger chat models, but embeddings quietly power a huge share of production AI: search, recommendations, deduplication, and retrieval-augmented generation. A small encoder that performs well on commodity hardware lowers the barrier for those pipelines.

• Compact footprint at about 350M parameters
• Bidirectional encoder architecture for embedding tasks
• Tuned for long-context inference on CPU
• Released under the permissive Apache 2.0 license

The Apache 2.0 license is notable in its own right, giving companies room to deploy and modify the model commercially. Full details and usage guidance are in Liquid AI's accompanying write-up.