Today we're releasing Orinth — an open-source AI studio for computer vision, NLP, and language models. Label, train, test, and serve, in one app, on your own hardware. Apache 2.0.
Pronounced OR-inth (/ˈɔːrɪnθ/).
The problem is not any single tool
The tools in this space are good. Roboflow is good at computer vision data. Label Studio is good at labeling. Weights & Biases is good at tracking. Axolotl and Unsloth are good at fine-tuning language models. LM Studio is good at chatting with one locally.
The problem is that a real project needs four of them, and they do not know about each other. You export a folder from one tool to import into the next. You lose provenance at every boundary. The metric you are reading in one tool cannot name the dataset version that produced it, because that version lives in a different product.
And every handoff is a place where "which model was this again?" becomes unanswerable.
What Orinth covers
| Modalities | Lifecycle | Runs where | Interface | |
|---|---|---|---|---|
| Roboflow | CV only | data → train | cloud | GUI |
| Label Studio | multi | label only | self-host | GUI |
| W&B | multi | tracking only | cloud | GUI |
| Axolotl / Unsloth | LLM only | fine-tune only | local | CLI |
| LM Studio | LLM only | chat only | local | GUI |
| Orinth | CV + NLP + LLM | label → train → test → serve | local, cloud optional | GUI + CLI |
One project holds your datasets, your runs, and the models promoted out of them. A dataset version is frozen before a run consumes it, and the run records which version it consumed. A scored model points at inference. A fine-tuned language model exports to GGUF and opens a chat.
Nothing gets exported to be re-imported, because there is no next tool.
Local is the default, not a mode
Orinth runs on your machine. There is no account, no telemetry by default, and no step in the core workflow that requires a network connection. Datasets, weights, predictions, and logs are written to local storage you can point at, back up, or delete.
Two things reach the network, both explicitly:
- Importing a dataset or base model from the Hugging Face Hub.
- The chat surface's web search toggle, while it is on.
We think this matters for two groups in particular. Small teams shouldn't have to upload a dataset to a vendor to draw a polygon on it. And in health, government, defense, and manufacturing QA, data residency is not a preference — it is the condition of the project existing at all.
What's in the box
- Dataset Studio — annotate segmentation masks, bounding boxes, and classification labels; split, preprocess, and freeze versions. Import from the Hub, or generate a dataset from your own documents with Data Recipes.
- Training — task-first runs over YOLO, Keras, and Hugging Face families, executing as subprocesses with streamed logs and epoch progress. Accelerator auto-detected.
- Testing — task-aware metrics and per-item inspection, with comparison runs that stack models side by side.
- Inference — single-shot predictions with overlays, confidence, timings, and full history.
- Chat — fine-tune with LoRA or QLoRA, export to GGUF, serve locally, and talk to it with sampler controls and a reasoning view.
- Architecture Studio — a visual graph editor with Keras and PyTorch emitters.
- Guided tours — every surface explains itself.
Where it's going
The open-source studio is the product, and it stays complete on its own. Two paid editions sit beside it for the cases local hardware can't cover: Orinth Cloud for GPU-backed runs and team workspaces, and Orinth Enterprise for SSO, RBAC, audit logs, and air-gapped deployment.
Neither will ever gate the local app.
Try it
Download the macOS build, or clone the repo and run make dev. If something is broken or missing, open an issue — this is early, and we would rather hear it now.
Orinth is a research and engineering instrument for measuring model behavior. It produces model outputs for human review — it does not provide clinical diagnosis or autonomous decision-making.