NEWVansa-3 is live: 4B parameters, 16k context

THE DECISION LAYER FOR YOUR SOFTWARE

Turn context
into decisions.

Vansa is a decision model API for the decisions inside your app. Define the possible answers, send your context, and get a category, a score, or a yes/no probability, with no generated text.

Start building What makes it different

Defined outcomes.No text generation.

CONTEXT → DECISIONExample from the docs

01 / INPUT STATEsupport_ticket.json

Duplicate charge on invoice #4411

“Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan.”

vansa-3 / System One

02 / STRUCTURED ANSWERSJSON

billing99.2%probability

billing99.22%

technical0.52%

sales0.09%

other0.17%

One state. Multiple questions. One API call.

Now running
Vansa-3.

4Bparameters

16ktokens of context per question

0generated output tokens

VANSA AT A GLANCE

Short answers.

Key facts about Vansa, in one place.

What is Vansa?: Vansa is a decision model API. You send a state (text or JSON) and typed questions (choice, score or yes/no), and it returns a calibrated probability for every option you defined, without generating text.
What's new in Vansa-3?: Vansa-3 is a new generation: about 4 billion parameters and a 16,384-token context per question, up from 2,048. It does better on questions it was never trained on (an estimated 0.871 against Vansa-2's 0.712 on 106 new questions), reasons better (80.1 against 44.3 on JevBench's public items) and its email answers are better calibrated. It is lower than Vansa-2 on jailbreak detection (0.915 against 0.975), slightly lower on email accuracy (96.1% against 96.4% on all 52 email questions) and lower on 11 of the 13 public classification tasks Vansa-2 was trained on. Requests that name vansa-1 or vansa-2 are answered by Vansa-3. What changed
How is it different from an LLM?: Chat models such as ChatGPT or Claude write text. Vansa-3 is built on a 4B-parameter language model, but it never writes: it reads the model's probability for each option you define, so every answer is one of your options, with a probability, and the same request gives the same answers.
Who is it for?: Developers who need classification, scoring or yes/no decisions inside their software: ticket routing, email triage, intent detection with up to 128 options, content checks, urgency scores and the next action in an agent loop. Link-building teams can use the trained email qualification preset for guest-post outreach threads.
How accurate is it?: On its trained email qualification preset, Vansa-3 agrees with the reference labels on 96.8% of answers across 26 questions, measured on 600 held-out threads (always picking the most common answer scores 81.8%). On 106 questions it was never trained on it scores an estimated 0.871 (Jev: 0.949). On the 13 public classification tasks Vansa-2 was trained on, Vansa-3 scores lower: 0.709 against 0.805 on average, and lower on 11 of the 13. Benchmarks
What does it cost?: Vansa is in closed beta: access is by private invitation only. API keys are issued by the Vansa team; there is no billing and no SLA yet.
Is it compatible with Jev?: Yes. POST /v1/systemone accepts the same state and questions format as TypeSafe's Jev, so Jev client code works after changing the base URL, the key and the model name.
What happens to my data?: Request bodies are not stored. Each request is logged with its time, key, model, question count, token count, latency, status and client IP, plus a short error message when a request is rejected (it can name the question id or model at fault), for quotas and support. Requests pass through Cloudflare and are answered on Vansa's own GPU server; no third-party AI provider receives them.
Which languages work?: Write in English where you can: Vansa-3's base model (JevK5) is documented by its author as English only. Its email skill was trained on threads that were 84% English, then Spanish and Portuguese; on held-out email threads (all 52 questions, unquantised run) the few Spanish (39) and Portuguese (11) ones scored about the same as the English ones (96.3% and 96.7% against 96.0%).
How do I start?: Read the quickstart, try requests in the playground, or point your AI coding agent at llms.txt.

01 / A DIFFERENT JOB FOR AI

Some AI writes the reply.
Vansa chooses the next step.

ChatGPT is a conversational assistant. Claude Code is a coding agent. Vansa is a decision model you call from your software.

CHAT & CODING ASSISTANTS

Create a response.

Use generative AI when you need a written reply, an explanation, or code. It creates content in response to your request.

TYPICAL TASK

“Draft a reply to this customer.”

A written response

VANSA / DECISION API

Choose from your outcomes.

Use Vansa when your app needs to classify, prioritize, or choose an action. It scores the answers you define and returns the result.

TYPICAL TASK

“Which department should handle this?”

billing99.2% probability

General language models can also classify inputs and return structured data. Vansa specializes in this decision step, without generating text. Vansa example shown from the docs ↗

Options you control.

Define the allowed categories, rating levels, or yes/no questions. The model stays within those choices.

Probabilities included.

See how the alternatives compare. Set a threshold for action or send uncertain cases for review.

Decisions in one call.

Check the category, urgency, and need for a reply together. Your code gets typed results, with no written answer to parse.

Use them together.Your application connects the steps.

INPUTA customer email
VANSARoute it to billing
GENERATIVE ASSISTANTDraft the reply

02 / THREE WAYS TO DECIDE

Your options.
Your next move.

From routing a support ticket to choosing an agent’s next action, get an answer in the shape your application needs.

choice

Pick the right option.

Define your categories. Get a selection and the probability of every option, with answers constrained to your set.

departmentbilling

Ticket routing · Intent detection

score

Put it on a scale.

Describe your ordered levels. Get an expected score and the distribution behind it, from low priority to critical.

urgency1.70 / 2

Urgency scoring · Sentiment analysis

noul

Ask a yes-or-no question.

Ask whether a statement holds. Get the probability of yes, so your application can act at the threshold you choose.

churn_risk85.68% yes

Email triage · Content checks

Example values: Vansa-3's real answers to the documented support-ticket request. Up to 32 questions per request; see API limits.

03 / STRAIGHT TO YOUR STACK

A little JSON.
A clear answer.

One endpoint for classification, scoring, and yes-or-no decisions. Use the tools you already work with.

01
Send your state
Text, a JSON object, or a conversation.
02
Define your questions
Set the instructions and allowed options.
03
Use the answers
Read typed results and probabilities.

Explore the API reference

curl https://api.vansa.org/v1/systemone \
  -H "Authorization: Bearer $VANSA_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "vansa-latest",
    "state": "Please refund my duplicate charge.",
    "questions": {
      "intent": {
        "type": "choice",
        "instructions": "What does the customer need?",
        "criteria": ["refund", "support", "other"]
      }
    }
  }'

Use the API key from your invitation. Keep it on your server.

CONTEXT IN. CLARITY OUT.

Make your next call.

Explore the API, or try Vansa with your own questions.

Open the playground

Closed beta: access is by private invitation from the Vansa team.

Turn context into decisions.