NVIDIA Nemotron 3 Ultra: Architecture and API Usage

Model Architecture and Performance

NVIDIA Nemotron 3 Ultra is a 550B parameter mixture-of-experts (MOE) model that utilizes a hybrid Transformer-Mamba architecture. While the total parameter count is 550B, the model activates approximately 55B parameters per token, balancing high-level reasoning with computational efficiency. NVIDIA claims the model is 5 times faster than competitors like Qwen 2.5 and GLM 5.1, while maintaining a 30% lower inference cost. While it excels at instruction following, it currently lags behind frontier models in long-horizon planning and agentic coding tasks.

API Implementation and Reasoning

The model is accessible via an OpenAI-compatible API endpoint. Developers can control reasoning depth through specific parameters in the chat completion request. Enabling the thinking process allows the model to output reasoning traces before the final answer.

# Enabling thinking and setting a reasoning budget
response = client.chat.completions.create(
    model="nvidia/nemotron-3-ultra-550b",
    messages=[{"role": "user", "content": "What is 2+2?"}],
    extra_body={
        "enable_thinking": True,
        "reasoning_budget": 1024
    }
)

For latency-sensitive applications, users can set low_effort: True to prioritize speed over reasoning depth. The model also supports native tool calling by passing a list of tools alongside the reasoning budget parameters.

Strategic Shift

NVIDIA is positioning itself as a major contributor to the open-weight model ecosystem. By releasing models across domains including speech (Parakeet, Canary), retrieval, and robotics (Groot), the company aims to accelerate the adoption of its hardware. This strategy functions as a flywheel where developing frontier models informs hardware design, while the resulting open-weight models drive demand for NVIDIA compute infrastructure.