NVIDIA NIM Provider Guide

August 8, 2026 ยท View on GitHub

VT Code supports NVIDIA NIM through NVIDIA's OpenAI-compatible Chat Completions endpoint.

Setup

Set an API key from the NVIDIA Build platform and select the provider:

export NVIDIA_API_KEY="nvapi-..."
vtcode --provider nvidia --model nvidia/nemotron-3-ultra-550b-a55b ask "Review this repository"

The default endpoint is https://integrate.api.nvidia.com/v1. Set NVIDIA_BASE_URL when routing through a compatible gateway or proxy.

Equivalent vtcode.toml configuration:

[agent]
provider = "nvidia"
default_model = "nvidia/nemotron-3-ultra-550b-a55b"

Curated models

These models appear in the NVIDIA section of the /model picker:

ModelContextNotes
nvidia/nemotron-3-ultra-550b-a55b1,000,000Default flagship agentic model
nvidia/nemotron-3-super-120b-a12b1,000,000Efficient long-context reasoning
nvidia/nemotron-3-nano-30b-a3b1,000,000Lower-cost reasoning and tool use
z-ai/glm-5.21,000,000GLM-5.2 served by NVIDIA NIM
deepseek-ai/deepseek-v4-flash-07311,000,000DeepSeek V4 Flash served by NVIDIA NIM

NVIDIA's catalog is larger than this curated list. An explicitly configured model ID is forwarded without local allowlist rejection:

vtcode --provider nvidia --model nvidia/any-catalog-model ask "Explain this code"

Reasoning and tools

VT Code maps any selected reasoning effort other than none to NVIDIA's chat_template_kwargs.enable_thinking = true; no effort or none disables thinking. When tools are present, VT Code also sends chat_template_kwargs.force_nonempty_content = true, as required by NVIDIA's Nemotron tool-calling format. Streaming usage totals and reasoning_content are preserved in both streaming and non-streaming responses.

References