vera-s1

Python package and Jev-compatible HTTP server for the vera System One decision models: vera-spark (~630M, compact), vera-core (~2B), vera-core-2 (~2B) and vera-large (~26B, most accurate, GPU only).

Give a vera model a state (text or JSON) and typed questions; it returns typed answers with calibrated probabilities in one forward pass. It never generates text.

pip install vera-s1            # add [fast] for flash-linear-attention (faster vera-core on GPU)

import vera

agent = vera.load("scar-ai/vera-spark")       # or "scar-ai/vera-core", "scar-ai/vera-core-2", "scar-ai/vera-large"
result = agent.predict(
    "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan.",
    {"department": {"type": "choice", "instructions": "Which department should handle this?",
                    "criteria": {"billing": "invoices, payments, refunds", "technical": "bugs, outages",
                                 "other": "everything else"}},
     "churn_risk": {"type": "noul", "instructions": "Does the user threaten to cancel or leave?"}})
print(result["answers"])

Question types: choice (2 to 255 options), score (2 to 10 ordered levels) and noul (yes/no probability).

HTTP server

vera-serve --model scar-ai/vera-core --port 8200
curl -s localhost:8200/v1/systemone -H 'Content-Type: application/json' -d '{
  "state": "I was charged twice. Please fix this ASAP.",
  "questions": {"billing": {"type": "noul", "instructions": "Is this ticket about billing?"}}}'

Same request and response shape as Jev's POST /v1/systemone. Binds 127.0.0.1 by default and has no authentication: put it behind your own gateway before exposing it.

By default a request is fitted to the model's trained budget: long option descriptions and questions are shortened and an oversized state keeps its head and tail. --no-caps (or VERA_NO_CAPS=1, or vera.load(..., no_caps=True)) reads every request whole instead. vera-core-2 and vera-large always read requests whole.

vLLM engine

vera-core, vera-core-2 and vera-large can run their backbone on vLLM (GPU, vLLM 0.29 or newer), behind the same HTTP API:

pip install "vera-s1[vllm]"
vera-serve --model scar-ai/vera-core-2 --engine vllm --port 8200
vera-serve --model scar-ai/vera-large --engine vllm --tensor-parallel-size 2 --port 8200   # or 1 on a 96 GB+ GPU

or in Python: vera.load("scar-ai/vera-core-2", engine="vllm"). vLLM runs the fine-tuned backbone and returns its hidden states; the decision head runs on top, so answers match the default engine up to bf16 differences. The backbone is exported to a plain Qwen3.5 checkpoint on first use and cached in ~/.cache/vera/vllm. A request longer than --max-model-len (default 32,768 tokens for vera-core-2 and vera-large, 8,192 for vera-core) is refused, never cut. --gpu-memory-utilization defaults to 0.5, or 0.9 for vera-large (~51 GB of bf16 weights). Plain vllm serve does not work on these repos: vera scores options and never generates text. vera-spark has no vLLM path.

Requires Python 3.10+ and transformers 5+. A GPU (CUDA or ROCm) is recommended; Apple MPS is not supported (CPU is used on Macs).