Skip to content

Integration

NVIDIA NIM logo

NVIDIA NIM

Evaluate, guard, and observe self-hosted models served through NVIDIA NIM inference microservices under the Hyperpriors control plane.

Request access to connect

Connections are provisioned in private beta — not authorised from this page.

01 — Permissions

When connected, Hyperpriors is limited to the following — declared up front, revocable at any time. Scope is granted during provisioning, not by a button on this page.

  • 01

    Route agent requests to self-hosted NIM endpoints through the Hyperpriors harness

  • 02

    Run evaluation suites against models served by NIM microservices

  • 03

    Enforce guardrail policies on NIM inputs and outputs at runtime

  • 04

    Record traces, token usage, cost, and latency for every NIM call

  • 05

    Track behavioural drift across NIM container and inference engine upgrades

02 — Details

Built by
NVIDIA
Website
build.nvidia.com (opens in new tab)
Category
Inference platform
Docs
docs.nvidia.com/nim (opens in new tab)

03 — Notes

About NVIDIA NIM

NVIDIA NIM is a set of inference microservices for running models on your own GPU infrastructure. Each NIM packages a model with an optimised inference engine in a prebuilt container that exposes industry-standard, OpenAI-compatible APIs, and can be deployed on premises or in any cloud with NVIDIA GPUs. NIM is distributed as part of NVIDIA AI Enterprise, with a catalogue of supported models available through build.nvidia.com. For teams that self-host for data residency, latency, or cost reasons, it removes much of the engineering work of standing up performant inference.

What the integration does

Hyperpriors treats a NIM endpoint like any other model target. The harness routes agent requests to your self-hosted endpoints with retries, timeouts, and fallback routes configured centrally, using the standard API surface NIM already exposes.

Self-hosted serving stacks change frequently: container releases, inference engine updates, quantisation changes, and GPU driver upgrades can each shift model behaviour without any change to the model you selected. Hyperpriors runs the same evaluation suites before and after each serving-stack upgrade and watches for behavioural drift between them, so a container update is a measured change rather than a silent one. Guardrails inspect inputs and outputs at runtime, and observability records every call with trace context, token usage, cost, and latency.

Get started

Point Hyperpriors at your NIM endpoints from the dashboard, or contact us to talk through your deployment.