Measured, then picked.
On NVIDIA's API, Ultra with its thinking off was the only setup to pass every test: 0 of 10 false rejections, 6 of 6 planted errors, 9 of 9 quotes verbatim. So it plans, screens, and backs up every judge.
NVIDIA Nemotronplanning, search, safety and every fallback
Every answer
runs through Nemotron.
Nemotron Embed finds every answer's passages and Nemotron 3 Ultra plans it; OpenAI then writes and judges. Lightning reads a spoken ask and plans deep searches, Ultra screens what they find, Content Safety guards what's shared, and when OpenAI is down or out, Ultra writes and checks instead.
ask three real papers The receipt under each answer names the models that made it.interactive examplenvidia × konkyo
Plans every answer
Nemotron 3 Ultra
thinking off
- its part
- first choice; gpt-6-luna if it's busy
- measured
- the only setup that passed every test on NVIDIA's API: valid JSON 3/3
- PlannerNemotron 3 Ultra
- ScreenerNemotron 3 Ultra
- ScoutNemotron 3.5 Lightning
- VoiceNemotron 3.5 Lightning
- WriterNemotron 3 Ultra (stand-in)
- VerifierNemotron 3 Ultra (stand-in)
Read from the app's own config: the Nemotron model in each job, whether it goes first or stands in, and what it was measured to do.
Follow one answer.
From one small action to what you see,
each step opened in the code.
- 1
Ultra turns it into searches.
One to three searches written in the words your passages would use, with your corrections read first; a stalled call is asked again, and OpenAI takes over if Ultra is busy.
- 2
Nemotron Embed finds the passages.
Your question is embedded as a query, your passages as passages, and pgvector and Postgres full-text search are merged by rank, so meaning and exact words both count.
- 3
OpenAI writes; Ultra is in reserve.
The Writer quotes one finding per sentence. If OpenAI fails, stalls or runs out of credit, Nemotron 3 Ultra writes it instead, and the switch is in the trace.
- 4
Content Safety guards what's shared.
A question asked in a public space, a public profile, and every article before it's published go through Nemotron Content Safety; a whole article is one check.
NVIDIA Nemotron in KonkyoWhere it
Where it
shows up.
Specific jobs in the app.
Each one has a place you can point to.
Deep search in bounded steps.
Lightning plans the searches; Ultra screens the papers and pages found eight at a time, carrying only the shortlist, so the context stays the same size however much comes back.
Asymmetric search.
Nemotron Embed reads a question and a passage differently, so each is embedded in its own role and cached under its own key.
Spoken asks, while you talk.
When no pattern places what you said, Lightning picks the action in about a second, with an 8-second deadline and OpenAI behind it.
A closer look.
The choices behind the behaviour.
Open whichever part you're curious about.
Asked twice, then a fallback
A Nemotron call that runs past its usual time is asked again and the first answer wins. One that fails, or gives nothing within 45 seconds (a stream: no first word within 25), runs on the other provider, and the switch is written to the run's trace.
Inside the free tier
NVIDIA allows about 40 requests a minute for each model. Each process keeps a sliding minute per model, the web app's and worker's shares add up to that, and a model that answers 429 goes second for a minute.
Only the failures are rewritten
Sentences the Verifier rejected go back to the Writer with its reasons; a later round judges only those, and every other sentence stays word for word.
1536 of 2048 dimensions
NVIDIA's API shortens Nemotron's vectors and normalises them again, which keeps them under pgvector's 2,000-dimension index limit.