Last updated: 10/13/2026Last verified: 2026-10-07
What is Vela 2.0?
Vela 2.0 is a family of Apache-2.0 open routing models from the vLLM Semantic Router project. It is designed to answer routing, safety, PII, and groundedness questions in a single request, returning decisions with probabilities and span-level character offsets. The models support request-time task definitions through choices, yes/no criteria, ordered rubrics, and label sets.
Key Features
- Four open models ranging from 0.3B to 9B parameters for encoder, hybrid, and flagship routing workloads
- Single SystemOne-compatible request can combine domain routing, safety, PII, and hallucination-related questions
- Span-level outputs identify PII, unsupported claims, and extractive evidence with character offsets
- Request-time task configuration using choices, yes/no criteria, ordered rubrics, and label sets
- Apache-2.0 licensing supports open use and adaptation, subject to the license terms
- Probability scores help developers inspect and tune routing or classification behavior
Best For
Pricing
Vela 2.0 is listed as free and open under the Apache-2.0 license based on the provided information. Users should still account for infrastructure, hosting, inference, and operational costs when running the models.
Pros & Cons
Pros
- Open model family with Apache-2.0 licensing
- Covers multiple routing-related tasks in one request
- Span-level character offsets improve interpretability for PII, evidence, and unsupported claims
- Multiple model sizes allow teams to choose between lighter and more capable deployments
- Flexible task definitions can be configured at request time
Cons
- Requires technical expertise to deploy and integrate into AI systems
- Operational costs may apply even though the models are free to use
- The provided information does not include benchmark details for specific workloads
- May require validation before use in production safety, compliance, or privacy workflows
- Documentation and ecosystem maturity should be reviewed by teams before adoption
Alternatives
Provides validation and guardrail tooling for LLM outputs, making it relevant for teams focused on safety and structured checks.
An open safety classification model family that can be used to evaluate content safety in LLM applications.
Offers an open-source framework for adding programmable guardrails to conversational AI systems.
Provides hosted content moderation models for teams that prefer an API-based approach to safety checks.
An open-source Microsoft tool for detecting and anonymizing personally identifiable information in text.
FAQ
Details
Platform
Features
- Four open models from 0.3B to 9B covering encoder, hybrid, and flagship routing workloads
- One SystemOne-compatible request combines domain routing, safety, PII, and hallucination questions
- Span-level answers locate PII, unsupported claims, and extractive evidence with character offsets
- Tasks are defined at request time through choices, yes/no criteria, ordered rubrics, and label sets
Languages
Known limitations
- Models output structured decisions only and generate no free-form text
- The 9B flagship needs roughly 32 GB of GPU memory in FP32 for evaluation
- The 0.3B tokenizer ships under Gemma Terms of Use instead of Apache-2.0
- Loading requires transformers 5.17 or newer with trust_remote_code enabled
