Get Started with GLM-5.3-Flash

This guide provides the model-specific deployment procedure and request examples for GLM-5.3-Flash. Before proceeding, complete the main Get Started flow.

Prepare FFmpeg 8

Video inference requires FFmpeg 8 and libavcodec.so.62. Obtain FFmpeg 8 from the FFmpeg 8.0 GitHub branch, and install the runtime in a directory on the host. Image and text inference do not require FFmpeg.

The examples on this page use $HOME/nim/ffmpeg8. Verify the runtime before starting the container:

export FFMPEG8_PATH="${FFMPEG8_PATH:-$HOME/nim/ffmpeg8}"
test -x "$FFMPEG8_PATH/bin/ffmpeg"
test -e "$FFMPEG8_PATH/lib/libavcodec.so.62"

Run NIM

Connected Deployment

Set the image, cache, and FFmpeg paths:

export NGC_API_KEY=<your-ngc-api-key> # optional for this NIM
export IMG_NAME=nvcr.io/nim/zai-org/glm-5.3-flash
export VERSION=2.1.2-variant
export LOCAL_NIM_CACHE="${LOCAL_NIM_CACHE:-$HOME/.cache/nim}"
export FFMPEG8_PATH="${FFMPEG8_PATH:-$HOME/nim/ffmpeg8}"
mkdir -p "$LOCAL_NIM_CACHE"

Start the container:

docker run -d --rm \
  --name="glm-53-$VERSION" \
  --gpus all \
  --shm-size=16g \
  -p 8000:8000 \
  -e NGC_API_KEY="$NGC_API_KEY" \
  -v "$LOCAL_NIM_CACHE:/opt/nim/.cache" \
  -v "$FFMPEG8_PATH:/opt/ffmpeg8:ro" \
  "$IMG_NAME:$VERSION"

For eligible public-catalog NIMs, you can omit NGC_API_KEY. For more information, refer to Installation.

Air-Gapped Deployment

Before starting the container in an air-gapped environment, transfer the container image, model files, and FFmpeg 8 runtime to the host. Set MODEL_PATH to the directory that contains the model files:

export IMG_NAME=nvcr.io/nim/zai-org/glm-5.3-flash
export VERSION=2.1.2-variant
export MODEL_PATH=/path/to/GLM-5.3-Flash
export FFMPEG8_PATH="${FFMPEG8_PATH:-$HOME/nim/ffmpeg8}"

Start the container with the local model directory:

docker run -d --rm \
  --name="glm-53-$VERSION" \
  --gpus all \
  --shm-size=16g \
  -p 8000:8000 \
  -v "$MODEL_PATH:/model:ro" \
  -e NIM_MODEL_PATH=/model \
  -v "$FFMPEG8_PATH:/opt/ffmpeg8:ro" \
  "$IMG_NAME:$VERSION"

The MODEL_PATH mount and NIM_MODEL_PATH setting are specific to the air-gapped deployment in this guide.

Verify the Deployment

After the container reports a ready state, query the models endpoint:

curl -s http://localhost:8000/v1/models

For a two-node deployment on DGX Spark, refer to Deploy on DGX Spark.

Control Reasoning Effort

Thinking is always enabled for GLM-5.3-Flash. Use reasoning_effort to control the reasoning budget. The default reasoning effort is max.

Reasoning Effort

Value	Behavior	Request Configuration
Max	Uses the largest reasoning budget and is the default.	"reasoning_effort": "max"
High	Balances reasoning depth and latency.	"reasoning_effort": "high"
Low	Uses a smaller reasoning budget for lower latency.	"reasoning_effort": "low"

For example, request low reasoning effort as follows:

curl -s http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3-Flash",
    "messages": [
      {
        "role": "user",
        "content": "Briefly explain why the sky appears blue."
      }
    ],
    "reasoning_effort": "low"
  }'

Send an Image Request

Send an image URL in an OpenAI-compatible chat completion request:

curl -s http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3-Flash",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "Describe this image."
          },
          {
            "type": "image_url",
            "image_url": {
              "url": "https://assets.ngc.nvidia.com/products/api-catalog/phi-3-5-vision/example1b.jpg"
            }
          }
        ]
      }
    ],
    "max_tokens": 1024
  }'

Send a Video Request

The container must have the FFmpeg 8 runtime mounted at /opt/ffmpeg8 before you send a video request.

curl -s http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "zai-org/GLM-5.3-Flash",
    "messages": [
      {
        "role": "user",
        "content": [
          {
            "type": "text",
            "text": "Describe what happens in this video."
          },
          {
            "type": "video_url",
            "video_url": {
              "url": "https://blogs.nvidia.com/wp-content/uploads/2023/04/nvidia-studio-itns-wk53-scene-in-omniverse-1280w.mp4"
            }
          }
        ]
      }
    ],
    "max_tokens": 1600
  }'