vera-large-2o

vera-large-2o is a non-autoregressive System 1 decision model that reads text and images. Give it a state (text, email, ticket, JSON, plus optional images) and typed questions; it returns typed answers with calibrated probabilities. It never generates text, so there is nothing to parse and nothing to hallucinate. It speaks the same wire format as Jev (POST /v1/systemone).

It is the second version of vera-large: the same Qwen3.8-27B decision architecture, now with the stock Qwen3.8 vision tower so the same checkpoint answers questions about images.

Benchmarks

We'll update this section once the official Jev Decision Index results for vera-large-2o are out.

Quickstart

vera-large-2o needs the vera package (vera-s1) 0.1.8 or newer:

pip install -U "vera-s1[fast]"          # add [vllm] for the vLLM engine

import vera

agent = vera.load("scar-ai/vera-large-2o")

state = "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan."
questions = {
    "department": {"type": "choice", "instructions": "Which department should handle this?",
                   "criteria": {"billing": "invoices, payments, refunds",
                                "technical": "bugs, outages, system errors",
                                "other": "everything else"}},
    "urgency": {"type": "score", "instructions": "How urgent is this?",
                "criteria": ["not urgent", "soon", "blocking"]},
    "churn_risk": {"type": "noul", "instructions": "Does the user threaten to cancel or leave?"},
}
print(agent.predict(state, questions)["answers"])

With images, pass them next to the state (file paths, data: URIs or PIL images in Python; data: URIs over HTTP):

vera-serve --model scar-ai/vera-large-2o --port 8200
curl -s localhost:8200/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "Receipt from a customer refund request.",
  "images": ["data:image/jpeg;base64,..."],
  "questions": {"total_ok": {"type": "noul", "instructions": "Does the receipt total match 42.50?"}}
}'

Question types: choice (2 to 255 options), score (2 to 10 ordered levels) and noul (yes/no probability). All questions in a call are answered together; requests are never truncated.

The transformers loader (AutoModel, trust_remote_code=True), the decision pipeline and the vLLM engine (vera-serve --engine vllm, or engines/ without the vera package) work as for vera-large, with images on every path:

from transformers import AutoModel
model = AutoModel.from_pretrained("scar-ai/vera-large-2o", trust_remote_code=True, dtype="bfloat16", device_map="cuda")
model.predict("Receipt from a customer refund request.", questions, images=["receipt.jpg"])

vera-serve --model scar-ai/vera-large-2o --engine vllm --port 8200    # vLLM runs the backbone and the vision tower

Limitations

• Needs a large GPU: ~51 GB of bf16 weights plus ~1 GB for the vision tower; an 80 GB-class card or larger, or two GPUs with vLLM tensor parallelism.