NVIDIA: Nemotron 3 Embed 1B (free)

NVIDIA Nemotron 3 Embed 1B is an open text embedding model from NVIDIA. It belongs to the Nemotron 3 family and is distributed as a free model on api14 under the identifier nvidia/nemotron-3-embed-1b:free. NVIDIA positions it as an embedding model rather than a generative chat model, so its job is to turn text into vectors that you can store, compare and search over. It was released by the provider on 2026-07-16.

The model takes text as its only input modality and produces embeddings as its only output modality. It has a context length of 32768 tokens, which is the amount of text you can pass in a single request, and a maximum output of 29491. Because the output is a vector representation rather than generated prose, there is no chat completion behaviour here: you send text in and get a numeric embedding back. The model is described as optimized for high-throughput, low-latency retrieval, which shapes where it fits best.

Those characteristics point to a fairly specific set of uses. Enterprise search is the obvious one: embed your document corpus once, embed the user query at request time, and rank by vector similarity. The same mechanism underpins retrieval augmented generation, where embeddings select the passages that a separate language model will later read. Code retrieval is another fit, since source files and natural language queries can both be embedded into the same space. Agentic retrieval also applies, where an agent needs to look things up quickly and repeatedly as it works through a task. In all of these cases the model is a component in a larger pipeline rather than the thing that writes the final answer.

You call it through api14's OpenAI-compatible API using the model id nvidia/nemotron-3-embed-1b:free. Because the endpoint follows the OpenAI embeddings shape, you send your text in the input field and receive a vector in response, so existing OpenAI client libraries and embedding code generally work with only the base URL and model name changed. The model lists no additional supported parameters, so you should not expect tuning knobs beyond what the standard embeddings request provides. Keep each request within the 32768 token context limit.

One practical consideration is that this is an embeddings model, so it will not generate text, hold a conversation or answer questions directly. If you need summarisation, reasoning or dialogue, you want a generative model instead, and this one is best used alongside it. The free tier also makes it a reasonable choice for prototyping a retrieval pipeline before you commit to a larger indexing effort.