AI model
nvidia/nemotron-3-ultra-550b-a55b
NVIDIA: Nemotron 3 Ultra is a 262,144-token model served on Composite at $0.500 per million tokens in and $2.20 per million tokens out. Ranked #9 of 179 models by traffic on Composite over the last 30 days, taking 1.5% of tokens and 1.05% of requests. Measured on Composite, 100.0% of its last 100 requests on Composite succeeded, the typical first token arrives in 1051 ms, output runs at about 26.7 tokens per second.
Try in Playground| Model ID | nvidia/nemotron-3-ultra-550b-a55b |
|---|---|
| Provider | DeepInfra |
| Context window | 262,144 tokens |
| Accepts | text |
| Returns | text |
| Input price | $0.500 per million tokens |
| Output price | $2.20 per million tokens |
| Prompt caching | Supported — cached input is billed at the provider's cache rate, on credits and plans alike |
| Thinking | Optional - see "NVIDIA: Nemotron 3 Ultra (Reasoning)" in the catalog to turn it on by default |
| Moderated | No |
| Cost on a monthly plan | About 5.0 plan credits per typical call (8k in, 1k out) |
| Routing endpoints | 2 (base plus 1 routing variant) |
Make a Composite API key and type nvidia/nemotron-3-ultra-550b-a55b as the model name. JanitorAI takes the full URL https://composite.lucidity.sh/v1/chat/completions; SillyTavern, ChubAI, Agnai and RisuAI take https://composite.lucidity.sh/v1. It runs on credits, a plan, or your own provider key.
JanitorAI setup · SillyTavern setup · Best Janitor AI proxies compared
#9 of 179
1.5% of tokens and 1.05% of requests over 30 days.
100.0%
Across the last 100 requests Composite served for this model.
1051 ms
Median across the same sample.
26.7 tok/s
Median generation rate after the first token.
7.8%
Share of prompt tokens served from the provider's cache over 7 days.
The same model behind 1 alternative endpoint. Call the id directly to pin a route; the base id above picks for you.
| Route | Model ID | Provider | Price |
|---|---|---|---|
| cheapThink | nvidia/nemotron-3-ultra-550b-a55b:cheapThink |
DeepInfra | $0.500 per million tokens input · $2.20 per million tokens output |
$0.500 per million tokens input · $2.20 per million tokens output
Credits pay the provider’s price plus 25%; cache hits cost less.
On a monthly plan, a typical call (8k tokens in, 1k out) uses about 5.0 plan credits, at the provider's own price.
DeepInfra
Nemotron models are tuned by NVIDIA for instruction-following and tool use, and tend to behave predictably in structured prompts.
With a 262,144-token context window, it can hold a long-running roleplay or a full lorebook in memory without losing earlier plot details, so it suits multi-session campaigns and character cards with heavy backstory.
It sits in the mid-range on price, a reasonable pick for regular roleplay use without going for the cheapest option available.
NVIDIA Nemotron 3 Ultra is an open frontier-reasoning and orchestration model from NVIDIA, with 55B active parameters out of 550B total (MoE). Built on a hybrid Transformer-Mamba mixture-of-experts architecture, it...
Supplied by DeepInfra, not written by Composite.
NVIDIA: Nemotron 3.5 Super VL · NVIDIA: Nemotron 3.5 Lightning · NVIDIA: Nemotron 3.5 Content Safety