---
license: apache-2.0
pipeline_tag: zero-shot-classification
base_model:
- vllm-sr/Decision-2.0-Lux-9B
base_model_relation: finetune
datasets:
- nvidia/Aegis-AI-Content-Safety-Dataset-2.0
- ToxicityPrompts/PolyGuardMix
- nvidia/Nemotron-Safety-Guard-Dataset-v3
- microsoft/llmail-inject-challenge
- OpenSafetyLab/Salad-Data
- KRLabsOrg/lettucedetect-prose-hallucination
- KRLabsOrg/lettucedetect-code-hallucination
- numind/NuNER
- urchade/pile-mistral-v0.1
- knowledgator/GLINER-multi-task-synthetic-data
- nlpaueb/finer-139
- tonytan48/Re-DocRED
- MultiCoNER/multiconer_v2
- martinjosifoski/SynthIE
- rajpurkar/squad_v2
- hotpotqa/hotpot_qa
- google-research-datasets/natural_questions
- KRLabsOrg/verbatim-spans
- KRLabsOrg/tool-output-extraction-swebench
language:
- ar
- zh
- cs
- nl
- en
- fr
- de
- hi
- it
- ja
- ko
- pl
- pt
- ru
- es
- sv
- th
tags:
- decision-model
- system-one
- qwen3.5
- semantic-router
- vela
- zero-shot-classification
- token-classification
- named-entity-recognition
- pii-detection
- hallucination-detection
- content-safety
- multilingual
- long-context
---
# Vela 2.0 9B
**Open Foundation Routing Models**
*Routing decisions. Safety checks. Precise text spans.*
The largest hybrid member of Vela 2.0: high-accuracy multilingual routing, long-document PII and hallucination detection, with open-label spans through one interface.
Define options, labels and rubrics at request time. Ask multiple named questions about a request, context and answer, and receive structured decisions with the text spans that support your workflow.
1. **High-accuracy decisions.** A 41.63 Jev Decision Index with optional Noul calibration, plus 0.989 AUC on unseen prompt-attack families.
2. **Routing and safety together.** Use one request for routing, prompt-attack checks, PII and unsupported-claim detection.
3. **Decisions at span resolution.** Return precise text locations for trained router labels and open-label extraction, with probabilities and character offsets.
| Specification | Value |
| --- | --- |
| Reported parameters | 7.9B |
| Backbone | 32-layer Qwen3.5 hybrid: Gated DeltaNet and gated GQA |
| Input limit | 16,384 tokens |
| Output | Choice, Yes/no (Noul), Score, Span and Set; no text generation |
| Evaluated precision | FP32 parameters; bf16 backbone autocast on GPU; FP32 heads |
| GPU parameter memory | About 32 GB in FP32 |
| Licence | Apache-2.0 |
## Quickstart
Install the dependencies, load the model and ask for routing and spans in one request:
For focused examples, see [Span questions](#span-questions-locate-text) and [Set questions](#set-questions-select-multiple-labels) below.
```bash
pip install torch "transformers>=5.17" safetensors tokenizers numpy
pip install flash-linear-attention # GPU: the Gated-DeltaNet kernel used in evaluation (optional, much faster)
```
```python
from transformers import AutoModel
m = AutoModel.from_pretrained("vllm-sr/Vela-2.0-9B", trust_remote_code=True)
m = m.to("cuda") # parameters load in FP32; on GPU the backbone runs under bf16 autocast (heads FP32)
PII_LABELS = m.vela2_engine.cal["pii_schema"]["labels"] # the 17 trained PII types
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"}},
})
```
Recorded response excerpt from this release; the [full response](USAGE.md#routing-and-span-questions) includes usage and thresholds:
```json
{
"answers": {
"domain": {
"type": "choice",
"choice": "health",
"confidence": 0.977,
"probabilities": {
"health": 0.988,
"math": 0.0,
"other": 0.012
}
}
},
"spans": {
"pii": [
{
"label": "PERSON",
"start": 8,
"end": 17,
"text": "Tom Baker",
"probability": 0.999
},
{
"label": "EMAIL_ADDRESS",
"start": 19,
"end": 40,
"text": "tom.baker@example.com",
"probability": 1.0
}
],
"halu": [
{
"label": "unsupported",
"start": 22,
"end": 29,
"text": "6 grams",
"probability": 0.982
}
]
},
"span_heads": {
"pii": "router",
"halu": "router"
}
}
```
GPU with bf16 autocast is the evaluated setting. FP32 on CPU or GPU is also supported; fp16 is not. Parameter memory above excludes runtime allocations. See [precision and execution](USAGE.md#precision-and-execution).
### 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:
```python
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"]`.
**Two span heads, one interface.** The router span head handles trained PII, hallucination and toxic labels. The broad span head handles open-label extraction. The engine selects a head for each span question; you can also set `"head": "router"` or `"head": "broad"`. See [span-head selection](USAGE.md#span-head-selection). For entities, relations and evidence examples, see [open-label extraction](USAGE.md#open-label-extraction).
### 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:
```python
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
| Question | Use it for | Output |
| --- | --- | --- |
| **Choice** | Route a request or select one of 2–255 supplied options. | Selected key and distribution |
| **Yes/no (Noul)** | Check a condition against the state. | P(yes) |
| **Score** | Rate against 2–10 ordered levels. | Expected level and distribution |
| **Span** | Locate personal information, unsupported claims, entities, relations or evidence. | Text spans, character offsets and probabilities |
| **Set** | Select any number of supplied labels. | Selected labels and per-label probabilities |
Span offsets refer to Unicode code points in the selected text. Span and Set answers also provide Yes/no views in the SystemOne response. [Output semantics and thresholds](USAGE.md#output-semantics) explain these fields.
## Results
Selected results for this checkpoint. [Full evaluation](EVALUATION.md) includes every benchmark, family comparison and scoring protocol.
| Benchmark | Result | Metric and mode |
| --- | ---: | --- |
| Safety, 14 public sets | **0.921** | Mean AUC; trained task families |
| Prompt attacks, unseen families | **0.989** | AUC |
| PII, 8K-token documents | **0.940** | F1; shipped calibration |
| Hallucination, 10,698 examples | **0.885** | Example-F1 |
| ACL-Verbatim evidence | **24.5** | Word-F1; broad span head, zero-shot |
| Jev Decision Index 0.2.1 | **41.63** | 38 benchmarks; `noul_calibration=True` |
- **Evaluation:** checkpoint selection, temperatures and thresholds used dev data. Safety results cover trained task families; they are not zero-shot comparisons.
- **Calibration:** Jev uses optional `noul_calibration=True`; the default score is 41.09. PII uses shipped calibration. [All evaluation modes](EVALUATION.md).
## Architecture
This model has its own 32-layer hybrid backbone, with Gated DeltaNet and gated GQA blocks, a candidate head, and separate router/broad span-head weights. [Editable SVG](assets/architecture/04-vela-2.0-9b-architecture.svg).
Operator and readout diagrams
**Gated DeltaNet**
Causal QKV convolution feeds the gated-delta update, followed by per-head normalization and output gating. [Editable SVG](assets/architecture/07-gated-deltanet.svg).
**Gated GQA and SwiGLU**
Gated GQA uses partial RoPE on Q/K and a sigmoid output gate; SwiGLU uses separate up and gate projections. [Editable SVG](assets/architecture/08-gated-gqa-swiglu.svg).
**Candidate readout**
CandidateHead combines scaled bilinear and additive MLP scores over runtime options, then applies task-specific calibration. [Editable SVG](assets/architecture/09-decoder-candidate-head.svg).
**Router and broad span heads**
Router and broad span heads have separate weights, with one selected for each span question. Overlapping word logits are combined before decoding. [Editable SVG](assets/architecture/10-decoder-span-heads.svg).
**State prefix and question isolation**
For a fixed rendered state, each question block sees the same state prefix and its own causal tokens. Additional span questions use separate sequences. [Editable SVG](assets/architecture/11-state-prefix-and-question-isolation.svg).
Choice, Yes/no, Score and Set share a candidate readout. Router and broad span heads use the same word-by-label topology with separate weights. The hybrid backbone reuses the state prefix within each rendered sequence; execution details are in [USAGE.md](USAGE.md#precision-and-execution).
## Reference
| Document | Contents |
| --- | --- |
| [USAGE.md](USAGE.md) | Full examples and recorded responses, typed parts, head selection, calibration, SDK and HTTP serving |
| [EVALUATION.md](EVALUATION.md) | Complete family results, research versus shipped PII modes, benchmark protocols and evaluation disclosures |
| [TRAINING.md](TRAINING.md) | Training stages, data mixtures, provenance, licences and release files |
| [PARITY.md](PARITY.md) | Export parity against the research scorer |
## Credit and citation
Vela 2.0 is led by KR Labs and [vLLM Semantic Router](https://github.com/vllm-project/semantic-router).
Read the [Vela 2.0 technical overview](https://vllm-sr.ai/blog/vela-2-0-open-foundation-routing-models/).
```bibtex
@misc{vela2_9b_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-2.0-9B},
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-2.0-Lux-9B](https://huggingface.co/vllm-sr/Decision-2.0-Lux-9B) (Apache-2.0), itself built on Qwen3.5-9B (Apache-2.0); their licences and notices are passed on in `LICENSE`, `NOTICE`, `ATTRIBUTIONS.md` and `LICENSES/`. Changes against Decision-2.0-Lux-9B: [MODIFICATIONS.md](MODIFICATIONS.md). Training data keep their own licences, some of them CC-BY-SA share-alike.
