Build a multimodal data pipeline that processes images, video, audio, and text through one system.
Run batch LLM inference or generate embeddings using OpenAI, Anthropic, or vLLM.
Scale a data pipeline from a single laptop to a distributed Ray cluster.
Benchmark multimodal AI workload performance against Ray Data or Daft.
| astrovela/vane | clayboxerreproduce/meccha-chameleon-tool | goldyubrain/videosubtitler | |
|---|---|---|---|
| Stars | 70 | 70 | 70 |
| Language | Python | Python | Python |
| Setup difficulty | moderate | easy | moderate |
| Complexity | 4/5 | 2/5 | 2/5 |
| Audience | developer | general | general |
Figures from each repo's GitHub metadata at analysis time.
Basic pip install is quick, distributed or GPU inference features need a Ray cluster and vLLM setup.
Vane is a data engine built for AI workloads that need to handle many kinds of data at once, such as images, video, audio, text, documents, sensor readings, and tables. It is built as a fork of DuckDB, an existing fast analytical database, and extends it so that different data types can move through the same processing system instead of needing separate tools for each one. The project currently consists of three planned parts. Vane Data, the piece that is actually available today, supports most of the features described in the README but is not yet considered production ready, and its interfaces may still change as the code is reviewed. Vane RL and Vane Agent are described as still in early design, with their source code planned for a future release. Vane Data lets you build data and AI pipelines using either DuckDB style SQL or a Python interface. It has built in support for calling large language models, generating embeddings, and running batch inference through OpenAI and Anthropic APIs, or directly through vLLM, a system for serving models efficiently. It also groups similar requests together to make better use of vLLM's prefix caching, which can speed up inference. Under the hood it tries to keep CPU, GPU, disk input and output, and model inference all running at the same time rather than waiting on each other, and the same pipeline can run on a single machine or scale out across a Ray cluster. The README includes benchmark results comparing Vane against two similar tools, Ray Data and Daft, on a single machine with 36 CPU cores, 64 gigabytes of memory, and one NVIDIA RTX 2080 Ti graphics card. These results are limited to a single node so far, and testing on multi node setups used by the original benchmark suite is still pending. To try it, install the vane-ai package from PyPI. It supports Python versions 3.10 through 3.14, with 3.12 recommended. The project is released under the Apache License 2.0 and credits DuckDB, Ray Data, Daft, and Trino as inspirations.
A multimodal data engine built on a DuckDB fork that lets you process images, video, audio, and text and run LLM inference at local or cluster scale.
Mainly Python. The stack also includes Python, DuckDB, Ray.
Free to use, modify, and distribute, including commercially, as long as you keep attribution notices, under the Apache License 2.0.
Setup difficulty is rated moderate, with roughly 30min to a first successful run.
Mainly developer.
This repo across BitVibe Labs
double-check against the repo, no cap.