Is this page helpful?

Current Limitations of the NIM Metadata API#

The following constraints apply to the NIM Metadata API today, so you can plan around them. For the fields these constraints refer to, refer to Response Fields.

Coverage#

Coverage is partial and growing. Metadata is attached at publish time, so NIM microservices released before this API launched in September 2026 generally lack metadata until they are republished. Design for a 404 status code as a normal outcome rather than an error.

Planned Field-Shape Changes#

The gpu and gpu_device fields can become arrays. A single profile can be valid on several compatible GPUs, which the current single-string fields cannot express.

To make that change a non-event for your integration, write your parser to accept either a string or a list now. For the pattern, refer to rule 4 in Notes for Automated Agents.

Unstandardized and Optional Fields#

The following table describes fields you should treat with care:

Field

Limitation

Guidance

throughput

Units and measurement conditions are not yet fixed, so figures are not comparable across NIM microservices.

Treat it as informational. Do not use it as a ranking key across NIM microservices.

workspace_hash

Optional today.

Do not assume its presence, but use it for integrity verification when it is present.

profile_id

Semantics vary by publication vintage, because identifiers generated by different versions of the publishing toolchain were computed differently.

Treat it as an opaque identifier. Compare it for equality, and never parse it or derive meaning from its structure.

Caching#

The API offers no ETag header, no Last-Modified header, and no conditional-request support. Cache based on tag stability instead, as described in Notes for Automated Agents.

Best Practices#

Apply the following practices to work within these constraints:

  • Treat a 404 status code as data, not failure. It is a definitive answer about metadata availability.

  • Parse permissively. Accept unrecognized keys and both shapes of the GPU fields.

  • Rank on fields with defined semantics. Use min_vram_per_device_gb and parallelism sizes rather than throughput when comparing NIM microservices.

  • Verify entitlement separately. An absent allowed_deployment_regions value means no recorded restriction, so confirm license terms rather than inferring them from absence.