Nemotron 3 Ultra

Nemotron 3 Ultra is a mixture-of-experts model with 550 billion total parameters but 55 billion active for each token. This design seeks to preserve the capacity of a very large model while activating only part of the network for an individual request, a useful approach for teams balancing complex work with inference efficiency. Third-party coverage positions it as a fast coding model intended to fit developer workflows rather than remain limited to conversational use.

The model is best suited to software-development and agent-oriented applications that benefit from strong reasoning, code generation, and high-throughput inference. Its reported performance includes more than 300 generated tokens per second and an Intelligence Index score of 48, while commentary from CodeRabbit highlights its relevance to coding benchmarks and real developer use. These results come from third-party sources, so teams evaluating production fit should test quality, latency, and deployment requirements in their own workloads.

CoreWeavenvidia/NVIDIA-Nemotron-3-Ultra-550B-A55Bnemotron

Open provider page Provider docs

Quick Info

Provider: CoreWeave
Model key: nvidia/NVIDIA-Nemotron-3-Ultra-550B-A55B
Release date: Jun 4, 2026
Last updated: Jun 4, 2026
Input modalities
Output modalities
Capabilities

Cost

Input token cost: $0.50
Output token cost: $2.15

Limits

Output tokens: 262,144 tokens
Context window: 262,144 tokens

Transparent token rates

Compare Nemotron 3 Ultra pricing

Rates are shown per one million tokens. Combined means one million input plus one million output tokens.

Browse this family

Latest news about Nemotron 3 Ultra

Weights & Biases

Coverage

Nemotron 3 is a family of three open-weight language models from NVIDIA released between December 2025 and June 2026: Nano (31.6B parameters, December 15, 2025), Super (120B parameters, March 11, 2026 at GTC), and Ultra (550B parameters, June 4, 2026 at Computex). All three share the same underlying hybrid architecture The family uses a hybrid Mamba-Transformer Mixture-of-Experts design with approximately 10:1 sparsity, activating roughly 10% of total parameters per token. Ultra, dropped to Hugging Face on June 4, 2026, is positioned as the reasoning engine of the lineup with over 300 tokens per second throughput, and the article cla

Weights & Biases

Coverage

Kilo Code announced that NVIDIA Nemotron 3 Ultra is available in its coding assistant for VS Code, JetBrains, and terminal workflows, free for a limited time. The model was introduced by Jensen Huang at Computex 2026 in Taipei and is described as the top open model on PinchBench with frontier-class intelligence and 5x Built on a hybrid Mamba-Transformer Mixture-of-Experts architecture, Nemotron 3 Ultra has 550 billion total parameters with only 55 billion activated per token, enabling over 300 tokens per second inference speeds. The model was explicitly engineered with contributions from the Nemotron Coalition for tool use, agentic

Weights & Biases

CoverageAnalysis

NVIDIA announced Nemotron 3 Ultra in Jensen Huang's Computex keynote, and Artificial Analysis evaluated the model in partnership with NVIDIA. The exact variant — 550B total parameters with 55B active (MoE), approximately 90% sparsity — is explicitly named, and the page confirms weights are offered in BF16 with NVFP4 qu On Artificial Analysis's Intelligence Index, Nemotron 3 Ultra scored 48, leading US open-weight models and ahead of Gemma 4 31B (39), Nemotron 3 Super (36), and gpt-oss-120b (33), though behind Kimi K2.6 (54). On a pre-release DeepInfra endpoint the model served over 300 tokens per second, well above the 50–100 tok/s t