We use cookies. Cookies and browser storage keep you signed in and remember your preferences. With your permission we also measure how the site is used. You can any time. See our cookie policy.
nvidia/nemotron-3.5-transcribeRuns on your balance at the prices above. New accounts start with free usage.
Settings
NVIDIA: Nemotron 3.5 Transcribe is an open-weight speech recognition model from NVIDIA. It is built as a 30B mixture-of-experts model with roughly 4B active parameters, which means only a fraction of the total weights are engaged per unit of work. It sits in the transcription category, so its purpose is narrow and specific rather than general purpose.
The model takes audio as its input and produces a transcription as its output. It handles recordings of up to one hour in length and supports 40 languages. No context window is defined for this model, which is consistent with transcription work: the unit of input is the audio recording itself rather than a token sequence you assemble. There is no text generation, no tool calling and no image or video handling here, so treat it purely as a speech-to-text component in a larger pipeline.
That shape of capability points to a fairly clear set of uses. Long-form recordings, such as meeting captures, interviews, lectures, podcasts and call recordings, are the natural fit, since the one hour ceiling covers most single-session audio without splitting. The 40-language coverage makes it usable for multilingual archives and for teams that process audio from more than one region. Because the output is a transcription, it slots in as the first stage of a workflow: transcribe, then pass the text to a separate model for summarisation, search indexing, translation, sentiment analysis or compliance review. It is also a reasonable choice when you want an open-weight recogniser you can reason about and host on your own terms rather than a closed hosted service.
You call it through api14's OpenAI-compatible API using the model id "nvidia/nemotron-3.5-transcribe". Because the interface follows the OpenAI convention, existing clients and SDKs work with a change of base URL and model id rather than a rewrite. The model exposes no configurable request parameters, so you send your audio and take the transcription as it comes back; there is no temperature, no top_p and no similar sampling control to tune. Plan your integration around a fixed request shape.
The main constraint to keep in mind is scope. This model transcribes and nothing else, so any task that needs reasoning over the text, generation, or non-audio input belongs to a different model downstream. If your audio exceeds one hour, you will need to segment it before sending. And if your work is text-only, a general language model is the better fit.
import fs from "node:fs";
import OpenAI from "openai";
const client = new OpenAI({
baseURL: "https://api14.com/api/v1",
apiKey: process.env.API14_API_KEY,
});
const result = await client.audio.transcriptions.create({
model: "nvidia/nemotron-3.5-transcribe",
file: fs.createReadStream("audio.mp3"),
});
console.log(result.text);
This community has no topics yet. Open the first one about NVIDIA: Nemotron 3.5 Transcribe.