Docs | Blog | GitHub | Vela 2.0 collection

Vela 2.0 0.3B

Open Foundation Routing Models

Routing decisions. Safety checks. Precise text spans.

Vela 2.0 brings routing questions and span-level decisions into one model interface. Supply your options, labels and rubrics at request time; ask about the request, its context and the answer together.

The 0.3B is the family's compact multilingual encoder, with an 8,192-token input budget and Torch or ONNX inference on CPU or GPU.

1. Route with your own criteria. Choose a route, check a policy condition, assign a score or select multiple labels.
2. Locate the text behind a signal. Return spans of personal information and unsupported claims with character offsets and probabilities.
3. Ask across the whole interaction. Use typed request, context and answer parts in one request, through Python or a SystemOne-compatible HTTP API.

Quickstart

Ask for personal information, unsupported answer text and a route in one request:

For focused examples, see Span questions and Set questions below.

pip install torch "transformers>=4.57" safetensors tokenizers numpy

from transformers import AutoModel

m = AutoModel.from_pretrained("vllm-sr/Vela-2.0-0.3B", trust_remote_code=True)
# Runs on CPU by default; optionally use m = m.to("cuda"). Keep Torch loading in fp32.
PII_LABELS = m.vela2_engine.cal["pii_schema"]["labels"]

result = m.system_one(
    state={"request": "Hi, I'm Tom Baker (tom.baker@example.com). What is the maximum daily dose of paracetamol for an adult?",
           "source": "For adults, the maximum dose of paracetamol is 4 grams in 24 hours, taken as 500 mg to 1 g every 4 to 6 hours.",
           "answer": "Adults can take up to 6 grams of paracetamol in 24 hours, in doses of 500 mg to 1 g every 4 to 6 hours."},
    questions={
        "pii": {"type": "span", "instructions": "Which spans are personal information?", "criteria": PII_LABELS,
                "over": "request"},
        "halu": {"type": "span", "instructions": "Which spans of the answer are not supported by the context?",
                 "criteria": {"unsupported": "a claim not supported by the context"}},   # over the answer by default
        "domain": {"type": "choice", "instructions": "Which subject area is this request about?", "over": "request",
                   "criteria": {"health": "medicine, clinical practice, nutrition, ageing or sexual health",
                                "math": "arithmetic, algebra, geometry, statistics or other mathematics",
                                "other": "a subject that fits none of the listed areas"}},
    })

Selected fields from the recorded Torch CPU response, with probabilities rounded to three decimals:

{
  "answers": {
    "pii": {"type": "noul", "noul": 1.0},
    "halu": {"type": "noul", "noul": 0.984},
    "domain": {"type": "choice", "choice": "health", "confidence": 0.51, "probabilities": {"health": 0.727, "math": 0.217, "other": 0.056}}
  },
  "spans": {
    "pii": [{"label": "PERSON", "start": 8, "end": 17, "text": "Tom Baker", "probability": 1.0}, {"label": "EMAIL_ADDRESS", "start": 19, "end": 40, "text": "tom.baker@example.com", "probability": 0.983}],
    "halu": [{"label": "unsupported", "start": 16, "end": 29, "text": "up to 6 grams", "probability": 0.895}]
  }
}

Offsets are Unicode code-point offsets into the targeted field. Full response, HTTP, SDK and ONNX examples.

Span questions: locate text

Use "type": "span" to extract text with labels and character offsets. Set criteria to a label-to-description mapping and over to the state field to inspect. This PII example reuses m and the trained PII_LABELS from above:

span_result = m.system_one(
    state={"request": "Hi, I'm Tom Baker (tom.baker@example.com)."},
    questions={
        "pii": {
            "type": "span",
            "instructions": "Which spans are personal information?",
            "criteria": PII_LABELS,
            "over": "request",
        }
    },
)
print(span_result["spans"]["pii"])

Each returned span includes label, text, start, end and probability. Offsets use Unicode code points, with start inclusive and end exclusive, in the selected field. The first example also shows hallucination detection: use the unsupported label, target over="answer", and supply grounding text in state["source"].

Set questions: select multiple labels

Use "type": "set" when zero, one or several labels can apply. Supply your labels and descriptions in criteria; each label is scored independently. This example reuses m from above:

set_result = m.system_one(
    state="My card was charged twice and the parcel never arrived.",
    questions={
        "issues": {
            "type": "set",
            "instructions": "Which issues does the customer report?",
            "criteria": {
                "billing": "payments, charges, refunds or invoices",
                "shipping": "delivery of an order or a parcel",
                "login": "signing in, passwords or account access",
            },
        }
    },
)
print(set_result["sets"]["issues"]["selected"])
print(set_result["sets"]["issues"]["probabilities"])

Read the selected labels and all per-label probabilities from sets["issues"]. Probabilities do not need to sum to one. Labels are selected when their probability exceeds the question's threshold; omit threshold to use shipped calibration, or add it to the question to set your own cutoff. The applied value is returned in set_result["thresholds"]["issues"].

Questions and outputs

All five types share one interface. Span and Set preserve SystemOne compatibility through additional response fields and Noul views. Request and response reference.

Results

• Selection: checkpoint selection used dev splits; final release selection also considered test results.
• Safety: these task families were part of Vela's training, unlike Decide's. This comparison measures router safety performance, rather than zero-shot generalisation.

Selected results above use identical rows for each task comparison. Full results and evaluation protocol include every benchmark, confidence intervals and the separate shipped, test-blind and research PII calibration paths.

Architecture

A 22-layer bidirectional ModernBERT encoder with schema-conditioned decision and word–label span readouts. The first attention pre-norm is Identity; this diagram shows the SDPA path. Editable SVG.

Operator and readout diagrams

Attention and GEGLU

The SDPA attention path uses 12 heads of width 64 and applies ×4 YaRN RoPE to Q/K. GEGLU has two 1,152-wide branches. Editable SVG.

Decision and span readouts

The decision and span readouts use 256-dimensional projections and shared readout LayerNorm parameters. Learned τ and inference calibration temperatures T are distinct. Editable SVG.

The backbone reads schema and typed state bidirectionally. Decision heads combine option cosine scores with an option MLP; the span head compares word and label states. Model, calibration and training details.

Reference

• Usage: Python, HTTP, SDK, Span/Set, typed parts, shortcuts and ONNX.
• Evaluation: full tables, scoring paths, references and selection disclosures.
• Training and model details: data, training recipe, input schema, calibration and repository files.
• Export parity: Torch/research-scorer and ONNX comparisons on the measured rows.

Credit and citation

Vela 2.0 is led by KR Labs and vLLM Semantic Router.

Read the Vela 2.0 technical overview.

@misc{vela2_unified_2026,
  title        = {Vela 2.0: Towards Open Foundation Routing Models},
  author       = {{KR Labs} and {vLLM Semantic Router}},
  year         = {2026},
  note         = {Blog post. Model: vllm-sr/Vela-2.0-0.3B},
  howpublished = {\url{https://vllm-sr.ai/blog/vela-2-0-open-foundation-routing-models/}}
}

Licence

Apache-2.0 for this model's weights, code and documentation. It is derived from Decision-1.0-Kai-0.6B (Apache-2.0) and Vela-1.0-Encoder-307M (MIT, from mmBERT); their licences and notices are passed on in LICENSE, NOTICE, NOTICE_SOURCES.json and LICENSES/. The tokenizer carries the Gemma Terms of Use: DISTRIBUTION_TERMS.md · License scope. Changes against Kai and Vela, including the eight added tokenizer entries: MODIFICATIONS.md. Training data keep their own licences, some of them CC-BY-SA share-alike (see NOTICE_SOURCES.json).

Explore the Vela model collection