Sandbox Models

Sandbox models run on VideoDB Sandbox Compute. Pick a model, create a sandbox on a compatible tier, wait until it is active, and pass the sandbox ID to a supported workflow. Sandbox models can be used across VideoDB workflows, including understanding analyzers, indexing pipelines, text-to-speech, image generation, audio generation, and text generation.

Use model names exactly as listed unless your VideoDB team contact gives you an alias.

Model catalog

Model	Type	Minimum tier	Good for
`google/gemma-4-E2B-it-FP8`	VLM	`small`	lightweight visual scene understanding
`Qwen/Qwen3.5-9B-FP8`	VLM / text generation	`small`	lightweight visual and text reasoning
`google/gemma-4-12B-it-FP8`	VLM	`medium`	balanced visual scene understanding
`google/gemma-4-26B-A4B-it-FP8`	VLM	`medium`	higher-quality visual scene understanding
`google/gemma-4-31B-it-FP8`	VLM	`medium`	advanced visual scene understanding
`Qwen/Qwen3.5-27B-FP8`	VLM / text generation	`medium`	larger visual and text reasoning
`openai/whisper-large-v3-turbo`	speech-to-text	`small`	transcription and speech recognition
`k2-fsa/OmniVoice`	text-to-speech	`small`	TTS, voice design, and voice cloning
`black-forest-labs/FLUX.1-dev`	image generation	`medium`	text-to-image generation
`stabilityai/stable-audio-open-1.0`	audio generation	`small`	text-to-audio and sound generation
`rtdetr-v2-r50vd`	object detection	`small`	object detection workflows

Choose a tier

Use the smallest tier that supports the largest model in your workflow.

from videodb import SandboxTier

# Small-tier models: OmniVoice, Whisper, Qwen 9B, lightweight VLMs, and object detection
small_sandbox = conn.create_sandbox(
    tier=SandboxTier.small,
    model_categories=["vlm", "text_to_speech", "speech_to_text", "object_detection"],
)

# Medium-tier models: FLUX and larger VLMs
medium_sandbox = conn.create_sandbox(
    tier=SandboxTier.medium,
    model_categories=["vlm", "image_generation"],
)

If a workflow uses both small and medium models, create a medium sandbox and reuse it across those jobs.

Where to pass the model name

Workflow	Model field	Sandbox field
Understanding analyzers	analyzer `config.model`	analyzer `config.sandbox_id`
Generation APIs	`model_name`	`sandbox_id`

For hosted/default models, omit sandbox_id. For models listed on this page, pass a compatible sandbox ID.

Use a model in Understanding and Indexing

Use the sandbox model during understanding, then index the produced artifact with the standard video.index(...) interface.

sandbox = conn.create_sandbox(
    tier=SandboxTier.medium,
    models=["google/gemma-4-31B-it-FP8"],
)
sandbox.wait_for_ready()

understanding = video.understand(
    analyzers=[
        {
            "type": "vlm",
            "name": "scene",
            "config": {
                "model": "google/gemma-4-31B-it-FP8",
                "sandbox_id": sandbox.id,
                "prompt": "Describe each scene in detail.",
                "schema": {
                    "scene_description": "string",
                    "activity": "string",
                    "setting": "string",
                },
            },
        }
    ],
)

understanding.wait_until_complete()
scene = understanding.get_analyzer("scene")

video.index(
    name="scene",
    source=scene,
    use_for=["semantic", "query"],
    fields={
        "semantic": ["scene_description", "activity", "setting"],
        "filter": ["activity", "setting"],
    },
)

For object detection, configure the model and sandbox in the analyzer, then choose filter and aggregation fields at indexing time:

understanding = video.understand(
    analyzers=[
        {
            "type": "object_detection",
            "name": "objects",
            "config": {
                "model": "rtdetr-v2-r50vd",
                "sandbox_id": sandbox.id,
            },
        }
    ],
)

understanding.wait_until_complete()
objects = understanding.get_analyzer("objects")

video.index(
    name="objects",
    source=objects,
    use_for=["query", "aggregate"],
    fields={
        "filter": ["object_labels", "object_counts"],
        "aggregate": ["object_labels"],
    },
)

Use a model in generation APIs

Generation APIs use model_name and sandbox_id.

job = coll.generate_image(
    prompt="A cinematic product photo of a smart camera",
    model_name="black-forest-labs/FLUX.1-dev",
    sandbox_id=sandbox.id,
)

image = job.wait(timeout=900, interval=5)

job = coll.generate_voice(
    text="Welcome to VideoDB Sandbox Compute.",
    model_name="k2-fsa/OmniVoice",
    sandbox_id=sandbox.id,
)

audio = job.wait(timeout=900, interval=5)

Pricing and limits

Sandbox pricing and concurrency limits are based on the sandbox tier, not the individual model. See Sandbox Compute for current pricing and limits.

Sandbox Compute

Create, use, and stop sandbox compute.

Understanding Artifacts

Configure analyzers and model selection.

Create an Index

Index artifacts produced by sandbox models.

Start Here

Core Concepts

Ingest

Understand

Act

Automate

Open Source Frameworks

Model catalog

Choose a tier

Where to pass the model name

Use a model in Understanding and Indexing

Use a model in generation APIs

Pricing and limits

Sandbox Compute

Understanding Artifacts

Create an Index

​Model catalog

​Choose a tier

​Where to pass the model name

​Use a model in Understanding and Indexing

​Use a model in generation APIs

​Pricing and limits

Sandbox Compute

Understanding Artifacts

Create an Index

Model catalog

Choose a tier

Where to pass the model name

Use a model in Understanding and Indexing

Use a model in generation APIs

Pricing and limits