NVIDIA Nemotron 3 Ultra NVFP4

by NVIDIA

NVIDIA Nemotron 3 Ultra NVFP4 is the 4-bit floating-point build of NVIDIA's largest open-weight reasoning model, pairing 550B total parameters with 55B active per token through a hybrid Mamba-Transformer LatentMoE architecture. It carries Multi-Token Prediction layers for native speculative decoding and is built for long-context, high-complexity agentic workloads. NVFP4 is NVIDIA's 4-bit floating-point format, and this build is pre-trained directly in NVFP4 using a quantization-aware recipe, keeping the bulk of its layers in that compact representation while holding select layers in higher precision for stability. The result is a frontier-scale Mixture-of-Experts that needs far less accelerator memory to run, supported on NVIDIA Hopper and Blackwell GPUs and deployable through frameworks like vLLM, SGLang, and TensorRT-LLM. The model targets the most demanding agentic AI, including complex multi-step agents, long-context analysis, and high-accuracy reasoning over code, math, and science. It is released under the permissive OpenMDW License (v1.1), alongside a BF16 build for different hardware configurations.

Key info

Input
Output
Features
Context window
262K
Max output
Input price
$0.60 /1M
Output price
$2.40 /1M
  • Zero data retention on pay-as-you-go
  • No training by default
  • no-KYC deposit available

Available routes

NVIDIA Nemotron 3 Ultra NVFP4 runs on 1 route through the qynio gateway. Compare residency, zero retention, and training posture at a glance — full data-handling detail per route below.

ProviderRegionZero data retentionTrainingInputOutput
MultiZero data retentionNo$0.60$2.40

Uptime and availability

NVIDIA Nemotron 3 Ultra NVFP4 runs on a single route through the qynio gateway today, so its availability is that provider's availability.

100%route uptime, last 30 days
Single route, no failover target within this model.

Name a second model in the same request and the gateway tries it on retriable errors, so a busy hour never has to reach your users. .

Measured over the last 30 days from each provider's official status feed via StatusGator. Refreshed hourly. See uptime for every provider qynio monitors.

Data handling per route

Each route hosting NVIDIA Nemotron 3 Ultra NVFP4 has its own privacy posture, residency, and no-KYC terms. Postures are maintained by qynio with a last-verification timestamp.

Fireworks Multi-region

Zero data retention is on by default on Pay-as-you-go — no action required. No training on customer data. GLOBAL; wallets; deposit available.

Zero data retention
On by default on Pay-as-you-go. Derived from the logging and moderation facts.
Training
No training on customer data.
Logging
None
Moderation
Not established
Caching
Not established
Subprocessor access
Subprocessors may read content
no-KYC deposit
deposit available
Transfer mechanism
wallets

Get started

Call NVIDIA Nemotron 3 Ultra NVFP4 through the qynio gateway with one API key. Let your coding agent set it up, or call it directly — qynio is drop-in compatible with the OpenAI, Anthropic, and Google AI SDKs.

Set it up with your agent

Copy this and paste it into a coding agent like Claude Code, Cursor or Codex and it'll wire up qynio for you.

Or call it directly

import OpenAI from "openai";
const client = new OpenAI({
apiKey: process.env.QYN_API_KEY,
baseURL: "https://api.qyn.io/v3/compat",
});
const completion = await client.chat.completions.create({
model: "fireworks/nemotron-3-ultra-nvfp4",
messages: [{ role: "user", content: "Hello" }],
});
console.log(completion.choices[0].message.content);

Compare NVIDIA Nemotron 3 Ultra NVFP4 with…

Side-by-side on privacy, crypto rails, pricing, and benchmarks.

Other models from NVIDIA

Start building with 700+ models

One API key. Every major provider. Up and running in minutes.

Get startedView Documentation