by NVIDIAReleased 4 June 2026
Nemotron 3 Ultra reads text and writes text back, thinks before it answers and can use the tools you connect. It is the biggest member of its family, meant for long-running assistant work such as coordinating other assistants, running coding assistants and deep research, where planning over many steps matters.
Reads about 510 pages · writes up to 460 pages
Standard OpenAI fields, grouped by what they control. Anything not listed is accepted and ignored.
Shaping the reply
How long, how varied, where to stop.
max_tokenstemperaturetop_ptop_kmin_pstopfrequency_penaltypresence_penaltyrepetition_penaltylogit_biasReasoning
Let it think before answering — and see the thinking.
reasoningreasoning_effortinclude_reasoningStructured output
Get JSON you can parse instead of prose.
response_formatstructured_outputsTools
Let the model call your functions.
toolstool_choiceNo. NVIDIA has retired it. Other NVIDIA models are on sale.
Nemotron 3 Ultra reads text and writes text back, thinks before it answers and can use the tools you connect. It is the biggest member of its family, meant for long-running assistant work such as coordinating other assistants, running coding assistants and deep research, where planning over many steps matters.
NVIDIA names no languages for Nemotron 3 Ultra (batch). Arabic, French and Darja: not stated.
It reads text. Up to 512K tokens in one request, about 510 pages. It writes up to 346 000 words at once.
Yes, its weights are public. Open weights on Hugging Face
NVIDIA. Released 4 June 2026.
Nemotron 3.5 ASR Streaming Multilingual 0.6B takes audio and writes out what is said, built to respond quickly in over 40 languages. It suits live work such as real-time captions and voice assistants, or pipelines that transcribe several languages, where text should appear while the speaker is still talking.
This model is designed to listen to audio recordings and convert the spoken words into written text. It excels at the task of accurately transcribing spoken content from various sources. A key feature is its ability to handle speech in 40 different languages. This model is strictly for speech recognition and cannot perform other general language or visual tasks.
NVIDIA's 600M multilingual speech recogniser trained on over 670,000 hours, detecting every official EU language by itself and returning punctuated text with segment timestamps at a 6.34% average word error rate
Nemotron 3.5 Super VL reads text, pictures, documents, screenshots of app screens and video, and writes text back, taking in about a thousand pages at once. It is made for assistants that look at a screen or a document, then think and use the tools you connect. It thinks unless you switch that off, and you can limit how much effort it spends thinking. Nemotron 3 Super is the text-only model of the same size.
From what its maker and two benchmark sites publish, and where it stands on OpenDunes. Read 7 Nov 2026.
Thinks first
Works through the problem before it answers.
Uses your tools
Can call functions your code provides.
Fills your JSON
Answers in the JSON shape you give it.
Searches the web No.
Can look things up online while it answers.
Cheaper on repeats
A prompt it saw a moment ago costs less the second time.
Open weights
The model's weights are public.