FOUR-SPARK BLACKWELL INFERENCE
Your private frontier
model cluster.
One model-aware endpoint backed by four DGX Sparks and a switched 200 GbE ConnectX-7 fabric. Live telemetry is public; inference requires an API key.
NOW SERVING
Discovering model…
Waiting for the inference backend.
https://dgx.cautela.io/v1
LIVE TELEMETRY
Performance at a glance
Not sampled yet
ACCELERATORS
Cluster load
Waiting for GPU telemetry…
LIFETIME
Serving counters
- Completed requests
- —
- Generated tokens
- —
- Prompt tokens
- —
- Prefix cache hit rate
- —
- Average first token
- —
- Preemptions
- —
GATEWAY
Request pressure
- Active requests
- —
- Streaming / buffered
- —
- Oldest active
- —
- Admission queue
- —
- Overload responses
- —
- Client disconnects
- —
- Maximum header wait
- —
- Last 256 ≥ 120 s
- —
INTRA-DGX FABRIC
Switch traffic
Bidirectional traffic
—
200 Gb/s each direction
Spark 01 · Head →—→ Spark 02 · Worker
— utilized
Spark 01 · Head ←—← Spark 02 · Worker
— utilized
Head link—
Worker link—
Spark 03 link—
Spark 04 link—
FEC corrected—
FEC uncorrected—
RX errors—
TX drops—
GET CONNECTED
Use the cluster from your tools
Ask the operator for a personal dgx_… key. Keep it in an environment variable and never commit it.
01
Launch Claude Code natively
Claude Code talks directly to the server's Anthropic-compatible Messages endpoint. No LiteLLM bridge is involved.
Loading current model…Requires the claude CLI. The model aliases ensure every Claude Code agent role uses the active DGX model.
02
Launch OpenCode
A temporary in-memory provider config points OpenCode at the OpenAI-compatible endpoint without writing your key to disk.
Loading current model…Requires the opencode CLI. Choose another active model above to regenerate the snippet.
03
Use any OpenAI-compatible client
Set the base URL to https://dgx.cautela.io/v1 and use your personal key as a Bearer token.
Loading current model…Streaming uses standard server-sent events. Anthropic clients may send the same key in x-api-key.