RETENTION
none
OpenAI-compatible inference on dedicated NVIDIA H200 hardware. Your prompts and completions are never written to disk by our stack — not to logs, not to traces, not to a database, at any hop.
Logging is metadata-only by construction against a whitelisted field set. The usage store has no content columns. Engine request logging is disabled at launch. See all 13 surfaces →
Hy3 from the official FP8 checkpoint at its complete 262K native window. The same precision every incumbent runs — no quiet quantization, no truncated context.
Every deploy clears a streamed tool-call matrix and a usage check proving reported tokens equal tokens actually generated.
Overload sheds as 429. Capacity limits and backend degradation are converted to 429s at the gateway within milliseconds, rather than surfacing as errors or an unbounded queue — routing sees backpressure, not failure.
Separated infrastructure. The gateway runs on different hardware from the GPU nodes, so a node swap is not an endpoint outage.
Nodes earn their traffic. Per-GPU GEMM straggler sweep, NVLink verification, ECC watch and a temperature-0 canary against committed baselines — before a node serves anything.
Engine images are pinned. No version bump reaches production without the tool-call matrix and the canary both green.
Any OpenAI-compatible SDK works unchanged — set the base URL and go. Write to [email protected] for a key.