Open source · Python · Open-weight models
Marigold is a package manager and runtime for local AI applications.
Every local AI tool brings its own model server, its own copy of the weights and its own claim on the GPU. Marigold puts one model layer under all of them. Applications install by name from registries you choose, share one model cache, and queue for the GPU. Chat and embeddings on OpenAI-compatible endpoints, image, speech and the other model types on a typed API -- on a laptop, a box at home, or your own cloud account, air-gapped once models are cached.
pip install bayis-marigold
marigold cache init
marigold registry add https://examples.marigold.run/
marigold application install chat
marigold cache populate chat
marigold application start chat
Applications run on infrastructure you control, from a laptop to your own AWS or Azure account, and no third-party model provider sits in the path. Where your data goes is a property of where Marigold runs. Data sovereignty →
01
One shared weight cache, independent of which application is running. Downloaded once, reused by everything that needs it. A new application costs a download of its code and whatever models it adds.
02
A marigold.toml, a models.yaml and the code that uses them, installed from a registry by name. The code runs in its own container and reaches models only through the API, never on the host.
03
An index, the archives and a catalogue, built with one command and served from GitHub Pages or any web server. Anyone can run one: for the public, for a team, or on an air-gapped network. Finding applications →
Why the split: models as a fat protocol, applications as thin clients on top -- the shape that made early internet protocols durable. Read the essay →
01
Everything under one directory: the bundled Postgres, a local GPU or CPU worker, outputs on disk. Several applications share it, and requests queue rather than fight for the GPU.
02
A managed Postgres such as Supabase, rented GPU instances as workers, outputs in an S3 bucket. The same applications, with more throughput: add a GPU host and nothing else changes.
03
AWS RDS or Azure Database for PostgreSQL, GPU instances and object storage, all in your own account. Nothing routed through a third party. Deployment →
Bay Information Systems, who build Marigold, help teams deploy it and build applications on it. Services →
Extract text, embed, classify, and gate restricted content before it reaches an index. Text and image embeddings in one pipeline enable cross-modal search.
Generate, score for safety, aesthetic quality, and prompt alignment. Regenerate automatically until thresholds are met or attempts run out.
Compare an observation image against a reference by structural embedding. On deviation, segment both, diff the masks, and describe the discrepancy in natural language.
Extract entities, classify, summarise, and convert to speech in multiple languages from one application run. Output: a summary plus audio per language.
Classify large volumes of unlabelled rows against a small labelled set. Low-confidence predictions go to an instruct model for explanation. No retraining.
Embed and compare images from multiple locations in parallel. On change, describe it and assemble a report, delivered as text and audio on a schedule.
Each of these is a set of small applications. The ones that recur are the basis of a standard library for Marigold applications.
Marigold, the self-hosted inference platform
(bayis-marigold on PyPI), is unrelated to the Marigold
depth-estimation model from ETH Zurich.
Install the CLI, add the examples registry, install an application, cache its models and run it.
Declare models, write the code, package it, and publish a registry from a GitHub repository.
Where Marigold and Ollama differ: model types, API surface, request handling, and measured throughput.
Managed Postgres, GPU workers and object storage in your own cloud account, with registries on your own network.
Open source. Install it, run it on your own hardware.
Questions? ed@bayis.co.uk