NVIDIA NIM Provider
VS Code extension that adds an NVIDIA NIM provider to Copilot Chat.
Requirements
- VS Code 1.104.0 or later
- GitHub Copilot extension installed and active
- An NVIDIA NIM API key from build.nvidia.com/models
Installation
From Source
- Clone this repository.
- Run
bun install --ignore-scripts && bun run compile.
- Press
F5 in VS Code to launch the Extension Development Host.
From VSIX
- Run
bun install --ignore-scripts && bun run package:vsix.
- Install the generated
.vsix file via the Extensions view (Install from VSIX...).
Setup
- Open Copilot Chat and choose the model picker.
- Select Manage Models, then add/configure NVIDIA NIM.
- Paste the API key obtained from build.nvidia.com/models.
- Select one of the NVIDIA NIM models returned by your account.
You can also run NVIDIA NIM: Manage NVIDIA NIM API Key from the Command Palette. The extension
will migrate that key into VS Code's language model provider group so the model picker can resolve
NVIDIA NIM models. The VS Code model settings flow is recommended for new setups.
Supported Models
The extension dynamically fetches available models from https://integrate.api.nvidia.com/v1/models.
It does not ship a hardcoded fallback model catalog; the Copilot Chat model picker shows the models
returned by your NVIDIA NIM account (DeepSeek V4 Flash, Kimi K3 / K2.6, Cosmos Reason2 8B, GLM 5.2, GPT-OSS, Nemotron 3,
Llama 3.3/3.2, Gemma 4, Muse Glimmer, Llama3 ChatQA, Palmyra, and more — 102 models live as of 2026-08-22).
Because the NVIDIA /models response currently omits capability metadata (context window, vision,
tool calling), the extension ships a static override table (src/model-catalog.ts) that fills in
display names, context windows and vision flags for known models, and applies conservative defaults
to the rest. Chat models are treated as tool-capable so they remain selectable in Copilot Chat Agent
mode; known non-tool VLMs (Fuyu, DePlot, Kosmos 2) are explicitly excluded from Agent mode, and
non-chat catalog entries (embedding, rerank, guard/safety, base/completion, translation models) are
filtered out of the picker. Non-chat model ids are re-checked whenever the catalog changes, so the
filter stays in sync with the live catalog (see docs/models.md for details).
Reasoning models (DeepSeek R1/V4, Kimi K3/K2.6, Cosmos Reason2, GLM 5.2, GPT-OSS, MiniMax M3, Nemotron 3, Nemotron
Super, QwQ) get a minimum output budget floor and silent retries so long thinking steps cannot
truncate the visible response.
Usage
- Open Copilot Chat (
Cmd/Ctrl + Alt + I).
- Select NVIDIA NIM from the provider selector.
- Choose one of the dynamically discovered NVIDIA NIM models and start chatting.
Development
bun install --ignore-scripts
bun run compile
bun run lint
bun run test -- --runInBand
bun run test:coverage
Press F5 in VS Code to launch the Extension Development Host.
Available Scripts
bun run compile – TypeScript コンパイルbun run watch – ファイル変更監視付きコンパイルbun run test – テスト実行bun run test:coverage – カバレッジ付きテスト実行bun run lint – ESLint チェックbun run lint:fix – ESLint 自動修正bun run format – Prettier フォーマットbun run package:vsix – VSIX パッケージ作成
Documentation
- Architecture — module map, data flow, retry and error handling
- Model Reference — model discovery, capability normalization, family adapters, and known quirks/workarounds
- Contributing — setup, testing, debugging, release steps
Marketplace Packaging
bun run package:vsix
The command above produces a .vsix that can be uploaded in the VS Code Marketplace publisher portal.
Privacy
- Your API key is stored securely through VS Code's language model provider configuration and, for
legacy command-palette setup, VS Code SecretStorage.
- Chat completions and model discovery requests are sent to
https://integrate.api.nvidia.com/v1.
