Vane

High-performance, multimodal-native engine for AI workloads.

Unifies multimodal data, intelligence, and continuous learning with Python and SQL interfaces, seamlessly scaling from local environments to Ray clusters.

Get Started Choose your workload
$ pip install vane-ai·Apache-2.0
import vane
vane.configure(runner="ray")
con = vane.connect()


assets = con.sql("""
    SELECT asset_id, uri, media_type
    FROM read_parquet('s3://raw-assets/*.parquet')
    WHERE media_type IN ('image', 'video', 'audio')
""")


features = assets.map_batches(
    DecodeAndInfer,  # user UDF; 1 model load/actor
    schema=feature_schema,  # explicit user schema
    gpus=1,
    actor_number=4,
)


features.write_parquet("s3://model-ready/features/")
Benchmarks

Multimodal inference benchmarks

Benchmarking multimodal AI pipelines across audio, video, document, and image workloads.

Vane Data vs Ray Data vs Daft · tuned elapsed time
Test environment

36 CPU cores · 64 GB RAM2080 Ti (modified VRAM) · 22 GB GPU memory

Document
Ray Data32.2%lowerDaft79.2%lower
Image
Ray Data35.1%lowerDaftOOM
Audio
Ray Data2.2%lowerDaftOOM
Video
Ray Data9.8%higherDaft8.6%lower
Full benchmarks
Elapsed time (seconds)
Tuned batch sizes · log scale · lower is better
Vane DataRay DataDaft
86.09127413
Document
1.15k1.77kOOM
Image
2.31k2.36kOOM
Audio
7.60k6.92k8.32k
Video

This single-node evaluation adapts the workload design from the Ray Data multimodal AI benchmark. Daft's OOMs on the image and audio workloads may reflect the test machine's limited memory, while the older 2080 Ti GPU also limits throughput. Given current hardware and compute-budget constraints, we have not reproduced the original benchmark's full cluster-scale setup; these results apply only to the recorded single-node environment.

Platform

A multimodal-native AI engine connecting data, models, and agents.

Unifies data, models, and agents — enabling continuous learning and scalable execution from local devices to Ray clusters.

Available now
1

Vane Data

Unified multimodal data from any source.

ImageVideoAudioTextDocumentsEventsSensorsTables
Coming soon
2

Vane RL

Train, evaluate, and align models across modalities.

RolloutTrajectoryRewardTrainingEvaluation
Coming soon
3

Vane Agent

Act in the real world, solve tasks, create value.

PlanningReasoningActionFeedbackMemory

FEEDBACK LOOP

New data, outcomes, and feedback continuously improve the system.

Vane Core

Local Runtime+Ray Runtime

Unified Multimodal Data Type

Sensors, metadata, lineage, and model artifacts under one execution semantics.

Streaming + Backpressure + Dynamic Batching

Continuous flow for large objects with adaptive batching and pressure control.

Overlapped Heterogeneous Execution

CPU, GPU, IO, and model inference overlap through asynchronous scheduling.

Edge-Cloud Coordination

The same pipeline runs across local devices and Ray clusters.

01

Install

pip install vane-ai

02

Run an example

Start from the docs examples and adapt the pipeline to your data.

03

Build your POC

Use the docs examples and llms.txt files to wire Vane into your stack.

Build your first AI pipeline on multimodal data.