Nemotron 3 Ultra
NVIDIA · released Jun 3, 2026 · nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
- Type
- Open weightsCustom licence
- Params
- 561B
- Context
- 512K
active per word not recorded by us · about 384K words of context · download allowed, licence restricts use
Our take
Written Sep 30, 2026Nemotron 3 Ultra is a large text model you can download, and its one clearly above-middle result is instruction following. Almost everywhere else it sits near the bottom of the boards, so it is a narrow tool rather than a general one.
Use it for constrained rewriting and format-following work, where it placed 14th of 58 on LiveBench Instruction Following as of 25 Jun 2026, or for long documents that go in whole without being split first. Its licence puts conditions on commercial use and redistribution, so read it before you build on it. Skip it if you need coding, data analysis or reasoning results near the top of a board, or an agent that recovers from failed commands and stays on task.
The case for it
- 14th of 58 on LiveBench Instruction Following as of 25 Jun 2026, a board of constrained-rewriting tasks, so it suits reformatting and rule-bound text work rather than open-ended problem solving.
- The request capacity takes long documents whole, though nothing here measures whether it recalls details from the middle of them.
- 30th of 55 on Arena Agent · Tool use as of 25 Sep 2026, mid-field and the only agent sub-board where it is not near the bottom.
The case against it
- 51st of 55 on Arena Agent as of 25 Sep 2026, with 52nd of 55 on Recovery and 49th of 55 on Task outcome, so autonomous multi-step work is not its job.
- 56th of 58 on LiveBench Data Analysis and 52nd of 58 on LiveBench Reasoning as of 25 Jun 2026, so coding, analysis and general reasoning are not where it earns its place.
- A custom licence puts conditions on commercial use and redistribution, so it needs reading before you build on it.
How good is it?
An open text model for chat and everyday questions, though it trails most models at multi-step tasks and changing course.
- multi-step work it carries out for youArena Agent · 51st of 55
- changing course when you give new instructionsArena Agent · Steerability · 49th of 55
- getting back on track after a step failsArena Agent · Recovery · 52nd of 55
EverydayGeneral questions and everyday reasoning
Not yet scored on Arena Text (overall). It is on LiveBench Mathematics, in 34th of 58 with 88.66.
CodingWriting and fixing code on its own
Not yet scored on Arena Coding. It is on LiveBench Coding, in 51st of 58 with 70.7.
AgenticPlanning, calling tools, staying on task
Arena Agent51st of 55 · −0.144
WritingDrafting and rewriting prose
Not yet scored on Arena Creative Writing. It is on LiveBench Language, in 54th of 58 with 70.81.
Placings on Arena's agent boards, from live sessions people ran themselves. A model can lead on one of these and sit mid-field on the others.
Boards this model appears on that none of the ratings above are built on.
Every published score for this model13 scoresEvery figure we hold, from 13 boards, with who ran it and a link to the source — including the boards no rating above is built on.
Can you run it yourself?
GeForce RTX 4090 · 24 GB
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Apple M1 Pro (16-core GPU) · 32 GB
Too large for this card. The weights do not fit even with part of them offloaded to system memory.
Apple M3 Ultra (80-core GPU) · 512 GB
Borderline fit on an estimated size. It leaves 17.9 GB spare on a size we calculated rather than measured, and a 10% error either way would change the answer.
Memory use by level
Against a 24 GB card.
What is quantisation? →This model on every device we track71 devicesThe Q4 build most people download, on each device: what the weights come to, how much context the memory leaves, and whether it runs. Smallest device that runs it first.
Check against your own machine → · Where to rent it hosted →
Or rent it from someone else
Prices checked between 4 hours and 10 hours ago — each listing carries its own date.
The only listing at 262K of context — the other 3 in the table below are not like-for-like. One cheaper row there is outside that comparison: a different quantisation.
- per 1M tokens
- $0.60 in / $2.40 out
- Context served
- 262K
- Throughput
- Not measured
| Provider | In / out per 1M tokens | Context | Throughput | Trains on prompts | Logs prompts | Zero retention |
|---|---|---|---|---|---|---|
| DeepInfrafp4Direct and through OpenRouter | $0.50 / $2.20checked 4 hours ago | 262K16K max reply through OpenRouter | 27 tok/sthrough OpenRouter | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterNo | DirectUnknownThrough OpenRouterConfirmed |
| OpenRouterOpenRouter's own listing | $0.60 / $2.40checked 4 hours ago | 262K | not measured | Unknown | Unknown | Unknown |
| Basetenfp4Through OpenRouter | $0.60 / $2.40checked 4 hours ago | 203K183K max reply | 55 tok/s | No | No | Confirmed |
| Venice AIfp8Through OpenRouter | $0.63 / $3.13checked 10 hours ago | 256K33K max reply | 12 tok/s | No | No | Confirmed |
DeepInfraAcross the 4 listings we hold: 3 say they do not train on prompts (1 of them only through OpenRouter), 0 say they do and 1 does not say. 3 appear in the zero-retention registry we check (1 of them only through OpenRouter); the rest are unknown to us.
What each host's API supports
From the parameter list each endpoint publishes. Streaming is omitted: nothing we hold reports it, for any model.
| Provider | Tool calling | JSON output | Strict schema |
|---|---|---|---|
| DeepInfrafp4Direct and through OpenRouter | ✓ | ✓ | ✓ |
| OpenRouterOpenRouter's own listing | ✓ | ✓ | ✓ |
| Basetenfp4Through OpenRouter | ✓ | ✗ | ✗ |
| Venice AIfp8Through OpenRouter | ✓ | ✓ | ✗ |
Tool calling: 4 of 4 listings say yes. JSON output: 3 of 4 listings say yes, 1 says no. Strict schema: 2 of 4 listings say yes, 2 say no.
When we formed this view
Recent changes
What moved
first indexed by our pipelineEach date is the day we first saw the change, or the day the maker announced it.
What we do not know about this model yet
- We hold no measured file for it, so all 3 sizes on this page are calculated from the parameter count.
- We do not hold the active parameter count for it, so how much of it runs on any one token is unknown to us.
- Nothing we hold says whether an endpoint streams, so we do not show it either way.
- 1 of 4 listings does not say whether it trains on prompts, and 1 answers only through OpenRouter, not for its own listing.
- We hold no batch or off-peak rate for any of its listings.
- We hold a decode speed for it, but no prompt-processing (prefill) figure, so how long the input side of a job takes is unknown to us.
Licence and identifiers
What the licence allowsCustom licence, what it allows commercially, and the identifiers you need to pull this model — its Hugging Face repo, our slug and a machine-readable card.
Licence
Custom licence
This model ships custom license terms that don't map to a known template. We haven't parsed them, so commercial use, redistribution and derivatives are unverified — review the original terms before shipping.
Identifiers
- Hugging Face
- nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B-BF16
- Architecture
- Mixture of experts
- Takes in, gives back
- Text in, text out
- Catalogue slug
- nvidia-nemotron-3-ultra