LiquidAI: LFM2.5-Embedding-350M (free)

LiquidAI publishes LFM2.5-Embedding-350M, a text embedding model from Liquid AI that is offered on api14 under the model id liquid/lfm-2.5-embedding-350m:free. It sits in the embeddings category rather than the chat category, so it is not a conversational model and does not generate prose replies. Its single job is to turn text into a vector that represents meaning, which makes it a building block for other systems rather than an endpoint users talk to directly. The provider released it on 2026-08-18.

The model accepts text as its only input and returns embeddings as its only output. It produces 1,024-dimensional vectors, a size that fits comfortably into most vector stores and similarity search libraries without special handling. Its context window is 512 tokens, and the maximum output is 460, which in embedding terms relates to the returned vector representation rather than generated text. Because the window is fairly small, inputs should be trimmed to the passage, sentence, or short document section you actually want represented, and longer material should be split before it reaches the model.

Those characteristics point to a few practical uses. Retrieval augmented generation is the obvious one: embed your document chunks once, embed the incoming question at query time, and pull back the nearest chunks by vector similarity. Semantic search over a product catalogue, support articles, or internal notes works the same way, since the model matches on meaning rather than exact keywords. Deduplication and clustering tasks also follow, where you group or collapse near-identical texts by comparing their vectors. Classification with a lightweight head on top of the embeddings is another option, as is reranking candidate results before sending them to a chat model.

Calling it through api14 means sending a request to the OpenAI-compatible endpoint with the model set to liquid/lfm-2.5-embedding-350m:free. If you already use an OpenAI client library, point its base URL at api14 and keep the same embeddings call shape. Note that the model lists no supported parameters, so extra tuning flags such as dimensions or encoding format adjustments are not available; send your input text and read the vector back from the response. Keep each request within the 512-token window to avoid truncation of the text you are trying to represent.

Because the context window is 512 tokens and no parameters are exposed, this model suits short, well-segmented text rather than whole documents or long conversations. If you need to embed large blocks in one pass, or you need control over vector size or other request options, a different embedding model will fit better. For straightforward sentence and paragraph level retrieval on short inputs, it covers the job.