NVIDIA Model Info Dashboard

Local dashboard for exploring the free models exposed through build.nvidia.com.

The app fetches the active model catalog, loads per-model metadata, flattens every metadata field into sortable table columns, and lets you probe live capabilities such as latency, context length, max output tokens, and tool calling support.

Highlights

• Shows only models that appear active and usable.
• Fetches model metadata for every listed model and renders it as a sortable table.
• Keeps the most useful columns pinned on the left: Live Ping, Model ID, Publisher, Context Limit, Max Output, Latency (ms), Tool Support, and Tested At.
• Supports global search, Exclude Inactive/Error, and Tool Support filtering.
• Probes live model behavior from the UI:
  • Ping re-tests one model.
  • Test Displayed Models tests displayed models that do not already have a complete live result.
  • Shift + Click on Test Displayed Models forces a full re-test of all displayed rows.
  • Backend probe requests are globally paced and automatically back off on 429 Too Many Requests.
  • Tool support probing tries multiple request variants, classifies explicit unsupported-tool responses, and retries accepted-but-truncated responses with a larger completion budget before giving up.
• Right-click any row to open a copyable cURL API example for that model.
• Force Refresh Data drops all saved test results, clears backend caches, and reloads the model list from NVIDIA with no cache reuse.

Quick Start

1. Install Node.js 18 or later.
2. Export your NVIDIA key in the shell:

export NVIDIA_API_KEY="your_nvidia_api_key"

3. Start the app:

./start.sh

4. The server starts on http://localhost:4920 by default and attempts to open the dashboard in your default browser.

Main Controls

Control	Behavior
Search	Filters rows by substring match across all visible values.
Exclude Inactive/Error	Hides rows whose live test state is Error or Inactive.
Tool Support	Shows only rows that have been tested and confirmed to support tool calling.
Ping	Re-tests one model and updates cached results.
Test Displayed Models	Tests displayed models that are still missing a complete live test result.
Shift + Click on Test Displayed Models	Forces a re-test of every displayed row.
Stop Testing	Cancels the running batch test.
Force Refresh Data	Clears all saved test data and backend cache, then fetches a fresh model list and metadata snapshot.

What The Live Test Actually Detects

Each live test can perform up to three NVIDIA API requests:

1. A small chat completion request to confirm availability and measure latency.
2. Metadata-aware token limit detection that prefers numeric metadata hints and falls back to an oversized max_tokens probe only when a live value is still missing.
3. An adaptive tool-calling probe that tries multiple compatible request shapes and can retry truncated accepted responses with a larger max_tokens value.

Tool Support is intentionally three-state:

• blank: not tested yet
• true: tool calling was observed
• false: the probe completed and concluded either that tool fields are explicitly unsupported or that accepted requests still never emitted tool calls

If NVIDIA rate-limits a probe, the row shows Rate Limited instead of being cached as a normal failure. Inconclusive tool support probes stay blank so they can be retried later. Hover the Tool Support cell to inspect the saved reason summary for false or inconclusive rows.

The right-click popover intentionally keeps only the hosted OpenAI-compatible cURL example. On 2026-04-14, https://integrate.api.nvidia.com/v1/messages returned 404, so the hosted endpoint used by this dashboard does not currently expose the Anthropic-compatible path that Claude Code requires.

Configuration

The runtime reads the API key only from NVIDIA_API_KEY. It does not use .env.

Optional backend environment variables:

• PORT default 4920
• MAX_CONCURRENCY default 12
• REQUEST_TIMEOUT_MS default 20000
• CACHE_TTL_MS default 300000
• PROBE_RATE_LIMIT_RPM default 36
• PROBE_MIN_INTERVAL_MS default derived from PROBE_RATE_LIMIT_RPM
• PROBE_TIMEOUT_MS default 15000
• TOOL_SUPPORT_TIMEOUT_MS default 25000
• PROBE_MAX_429_RETRIES default 2
• PROBE_429_BACKOFF_MS default 10000

Repository Docs

• USAGE.md
• REQUIREMENTS.md
• IMPLEMENTATION.md
• TESTING.md