pi-extension-nvidia-nim

NVIDIA NIM provider for the pi coding agent — access 100+ models hosted on NVIDIA's inference microservice platform, including DeepSeek, Llama Nemotron, Qwen, GLM, Mistral, MiniMax, and more.

Registers the nvidia-nim provider with pi, backed by https://integrate.api.nvidia.com/v1.

Features

• ~81 curated models — chat, reasoning, code, and vision
• 121 scraped entries filtered, deduplicated, and family-mapped
• 8 handler-based thinking formats — DeepSeek V4, DeepSeek NIM, thinking-budget, Nemotron system modes (3 variants), MiniMax inline, and Qwen chat-template — plus native pi handling for reasoning-effort
• Model-specific quirks handled automatically — per-model chat_template_kwargs injection (thinking effort, budgets, system-message toggles), content array normalization for older models
• No custom streaming — uses pi's built-in openai-completions

Install

pi install npm:pi-extension-nvidia-nim

Configure

1. Get an API key

Sign up at build.nvidia.com (free tier: 40 requests/minute, 1,000 inference credits on signup, no credit card required).

2. Set the credential (pick one)

Option A — Environment variable:

export NVIDIA_NIM_API_KEY="nvapi-..."

NVIDIA_API_KEY is accepted as a fallback for backward compatibility with pi's built-in nvidia provider.

Option B — Auth file (~/.pi/agent/auth.json):

Add an entry so pi resolves the key automatically for all NVIDIA providers:

{
  "nvidia": { "type": "api_key", "key": "nvapi-..." }
}

Option C — Interactive login:

Run /login in pi's interactive mode and select NVIDIA from the list. The key is stored in auth.json and managed automatically.

3. Run pi with the extension

Install from npm:

pi install npm:pi-extension-nvidia-nim

Usage

pi
/model
# or Ctrl+P to pick a model

Look for the nvidia-nim/ prefix in the model picker.

Design

• Uses pi's built-in openai-completions streaming — no custom streamSimple.
• Model-specific quirks (thinking formats, extra body kwargs, compat flags) are handled via before_provider_request and pi's compat system.
• Family-based config in config/model-families.ts (46 families, first-match-wins) drives thinking format routing and model metadata.
• All cost fields are $0 — NVIDIA NIM free tier.
• Coexists with pi's built-in nvidia provider — use nvidia-nim/... for the full experience, nvidia/... as a basic fallback.

Comparison with pi's built-in nvidia provider

Pi ships a built-in nvidia provider with ~20 models. This extension (nvidia-nim) complements it with a wider model selection and thinking/reasoning support:

| Aspect | Built-in nvidia | This extension nvidia-nim | |--------|-------------------|-----------------------------| | Models | ~20 curated | ~81 curated (full NIM catalog) | | Thinking formats | None | 8 handler-based formats + reasoning-effort | | Content normalization | No | Yes | | Rate-limit warnings | No | Yes (429 handler) | | API key | NVIDIA_API_KEY env | NVIDIA_NIM_API_KEY + NVIDIA_API_KEY fallback |

Both can coexist. Use nvidia-nim/... for the full feature set, nvidia/... as a lightweight fallback.

Models with thinking support

DeepSeek V4, Kimi K2.6, Qwen3, GLM-5.2, MiniMax M3, Seed OSS, Nemotron (Ultra, Super, 3-Super), GPT-OSS, and StepFun.

Notable:

• GLM-5.2 — full reasoning effort control (high/max) via enable_thinking and clear_thinking kwargs
• MiniMax M3 — three-mode thinking toggle (disabled/adaptive/enabled) mapped from pi's thinking levels
• Nemotron — system-message-driven thinking modes (detailed think, /think, and reasoning budget variants)
• DeepSeek V4 — reasoning_effort inside chat_template_kwargs with off→none and xhigh→max mapping

Additional capabilities

• Rate-limit warnings — surfaces HTTP 429 responses with retry-after info
• Content array normalization — converts [{type:"text"}] to plain strings for older models that reject structured content arrays
• 46-family regex routing — accurate thinking format and compat assignment across all ~81 models
• Per-model reasoning effort mapping — non-standard effort values are handled automatically (e.g. off→none, minimal→low)
• Architecturally clean — uses before_provider_request event hook with no custom streamSimple, avoiding provider conflicts