Vela model catalog
Vela 1.0 is the model family for intelligent routing. Its twelve releases share the 307M-parameter Vela Encoder foundation and cover request understanding, safety, retrieval, and reranking.
Choose a model
| Model | Use it to |
|---|---|
| Encoder | Train a new task on the shared foundation |
| Domain | Classify requests into 14 subject areas |
| Guard | Detect prompt injection and jailbreak attacks |
| Safety | Detect unsafe content |
| Shield | Detect unsafe requests; an alternative to Safety |
| Hazard | Identify 12 content-risk categories |
| PII | Locate 17 types of personal information |
| FactCheck | Decide when an answer needs factual verification |
| Feedback | Recognize satisfaction, clarification, correction, revision, or no feedback |
| Modality | Choose text, image, or combined output |
| Embedding | Compare requests and retrieve relevant documents |
| Reranker | Reorder retrieved candidates by relevance |
Guard detects attempts to redirect instructions. Safety detects content risk;
Hazard identifies its category. Use these models together when your application
needs both prompt-attack protection and content policies. Shield is a separately
trained model with the same safe/unsafe interface as Safety. Both are
explicit specialists; the default decision deployment is Vela 2.0 0.3B, and the Safety signal guide
shows how to select Shield instead.
Embedding and Reranker offer four encoder depths and five dimensions so you can balance quality, latency, and memory. See embedding and reranking for how to choose.
Deploy or customize
To run the published models, follow Use Vela models. That guide covers default downloads and router configuration. Model cards provide standalone quickstarts and evaluation results.
To adapt a model to your own data, start with the training overview. The shared Encoder is a training foundation; use a task checkpoint for a routing signal.
Earlier releases and multimodal models
The earlier mmBERT classifier collection and mmBERT embedding collection remain available for existing integrations and comparisons. Their original training entrypoints are listed in the artifact index.
Vela 1.0 contains text models. Image and audio embedding models have their own multimodal training guide.