Independently benchmarked · Artificial Analysis

The fastest Nemotron 3 Ultra.
Not by a little.

NVIDIA's frontier open-weights reasoning model — 550B parameters, served at 454 tokens per second. Ranked #1 for output speed among every API provider Artificial Analysis benchmarks, 47% ahead of the runner-up, at roughly a third of its price.

  • 454 t/s output
  • +47% vs #2
  • $0.44 / 1M tokens
  • Open weights · E2EE

Head to head · #1 vs #2

Same model. Same prompt.
Watch the gap open.

Identical weights, identical prompt, streamed side by side against Nebius — the #2 provider on the leaderboard. BLACKBOX AI answers sooner and streams 47% faster, so the same completion simply lands first. The readouts are the published Artificial Analysis figures.

Nemotron 3 Ultra · Nebius vs BLACKBOX AI
Output TPS0-31.9%
TTFA+55.4%
$ / 1M+172.7%
Output TPS0+46.9%
TTFA-35.7%
$ / 1M-63.3%
Nebius · #2
reasoning…
BLACKBOX AI · #1won
reasoning…
Side-by-side speed test: the identical prompt streamed through Nebius and through BLACKBOX AI, both serving NVIDIA Nemotron 3 Ultra. BLACKBOX AI streams 47 percent faster and finishes first.
+47% output t/sSame open weights$0.44 vs $1.20 / 1M

The leaderboard

Every provider.
One clear #1.

Artificial Analysis continuously benchmarks every API provider serving Nemotron 3 Ultra — same methodology, same token counting, published in the open. BLACKBOX AI holds the top output speed with the lowest time-to-first-answer among the fast providers, and one of the lowest prices on the board.

Verify the numbers · Artificial Analysis
nemotron 3 ultra · 6 providers[ LEADERBOARD ]

Output speed · ranked

t/s · TTFA · $ / 1M
#1BLACKBOX AI454.4$0.44
#2Nebius309.3$1.20
#3CoreWeave249.6$0.53
#4Together AI159.5$0.62
#5DeepInfra99.2$0.39
#6Lightning AI72.8$0.49
TTFA = time to first answer token · blended price per 1M tokensSource: Artificial Analysis · July 2026

The engine

Same weights as everyone else.
A very different engine.

01

Custom CUDA kernels

We generate tokens, not resell them. Direct access to open weights lets us self-host on our own infrastructure, with kernels tuned for exactly this architecture.

02

FP4 on NVIDIA B300

Advanced quantization on latest-generation hardware. On the same bench, FP4 leads BF16 on every axis — with no measurable quality loss on evals.

03

Multi-token prediction

Nemotron ships a built-in MTP head. Our engine exploits it for speculative decoding, emitting multiple tokens per forward pass instead of one.

04

Encrypted end to end

Speed without the trade-off: prompts and completions are encrypted in transit and at the edge, never logged, retained, or used for training.

Drop-in

Point your SDK at it.
Nothing else changes.

The BLACKBOX API is OpenAI-compatible — swap the base URL and key, set the model to blackboxai/nvidia/nemotron-3-ultra, and every token arrives 47% sooner. Streaming, function calling, and JSON mode work unchanged.

nemotron.py
python
from openai import OpenAI

client = OpenAI(
    base_url="https://api.blackbox.ai/v1",
    api_key="YOUR_BLACKBOX_API_KEY",
)

stream = client.chat.completions.create(
    model="blackboxai/nvidia/nemotron-3-ultra",
    messages=[{"role": "user", "content": "Explain CUDA, step by step."}],
    stream=True,
)

for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")