> This page is for version 1.0.0 (Latest) (default).
> For other versions, use one of these documentation indexes:
> - 1.0.0 (Latest) (default): https://docs.nvidia.com/nim/harmonizer/1.0.0/llms.txt

> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.nvidia.com/nim/harmonizer/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.nvidia.com/nim/harmonizer/_mcp/server.

## Release 1.0.0

These notes describe the first generally available release of Harmonizer NIM.

This is the initial release of Harmonizer NIM.

### Highlights

* Synchronous image and temporal processing over HTTP or gRPC.
* Independent-frame processing and rolling temporal context from the same
  Harmonizer checkpoint in bfloat16 (BF16) precision.
* Bring Your Own Checkpoint (BYOC) for compatible Harmonizer source `.pkl`
  checkpoints.
* GPU-specific executor specializations and conservative fallback
  specializations for unlisted GPU SKUs.

### Evaluation Notes

Quality and performance results are comparable only when they use the same
dataset, input resolution, processing mode, return policy, transport, codec, and
evaluator version. Model-card values or results from a different protocol are
not a matched comparison with NIM quality-gate results.

### All Known Issues

The following known issues and limitations apply to Release 1.0.0:

* Harmonizer runs inference at `1024x576`. Inputs at other resolutions are
  resized directly to the model resolution and outputs are resized back to the
  input dimensions. The resize does not preserve aspect ratio, letterbox, or
  crop; use 16:9 inputs to avoid geometric distortion.
* `context_length` supports only `0` and `4`. In temporal mode, the first
  four frames are processed independently while the rolling context is
  established. Later frames use the four previously generated outputs as
  temporal context.
* Inference is synchronous. Select either HTTP or gRPC inference when the
  container starts; both inference transports are not served simultaneously.
  HTTP health endpoints remain available in either mode.
* Each container uses one GPU. The generic fallback profile lets an unlisted GPU
  start with conservative executor specializations, but it is not a performance
  or support qualification for GPUs outside the [Support Matrix](/nim/harmonizer/support-matrix).
* Cold startup takes several minutes because model loading, compilation,
  CUDA-graph capture, and warmup finish before the service reports ready.
  Persisting the NIM cache can shorten later starts, but these per-process startup
  phases still run.
* BYOC accepts a complete, compatible Harmonizer source `.pkl` checkpoint.
  Adapter-only artifacts, training deltas, and checkpoints for other model
  architectures are not supported.
* NFS-backed bind mounts can reject cache access because of server-side identity
  mapping, ACLs, or stale file handles. Use a local-disk cache or run the
  container with a user that can write to the mounted directory.

### Related Topics

Continue with these related guides:

* [Quickstart Guide](/nim/harmonizer/quickstart-guide)
* [Troubleshooting Harmonizer NIM](/nim/harmonizer/troubleshooting)