Revyn Reflex
A family of small System One models for sales and support calls: what they are, why they exist, and what is available today.
Revyn Reflex is a family of small models that make the per-turn decisions a live call depends on. They do not write text. Each one reads a piece of a conversation and returns typed answers with a probability attached, fast enough to sit on the audio path.
The first model in the family is revyn-turn, which is in preview.
#Why a separate family
A voice agent, or a practice call with an AI buyer, runs on two kinds of work.
| System Two | System One | |
|---|---|---|
| The job | Compose a reply, reason about a deal, write a debrief | Decide what just happened in this turn |
| Typical output | Paragraphs | One label from a fixed set |
| Model | A large language model | A small task-specific model |
| Budget per turn | Hundreds of milliseconds to seconds | Tens of milliseconds |
Most stacks use the same large model for both. That works, but it means a full generation request to answer "is this buyer asking for a human?", on every turn of every call. The alternative most teams fall back to is a keyword list, which is fast and breaks as soon as a buyer phrases it the way buyers actually do, especially in code-mixed speech.
Reflex models take the System One half. The language model stays in the loop as System Two, and as the fallback whenever a Reflex model is not confident.
#How a Reflex model answers
Every question a Reflex model answers is typed, and the answer set is fixed in advance.
| Type | Shape | Example |
|---|---|---|
| Choice | One of N named options | Which objection is the buyer raising? |
| Graded score | One of a few ordered levels | How frustrated is the buyer? |
| Yes or no | Two options | Is the buyer asking for a human? |
For each question the model returns the chosen answer, a probability for every option, and a confidence. The probabilities are calibrated after training, so a confidence of 0.9 is meant to be right about nine times in ten. That is what lets calling code set a threshold: act on the answer above it, hand the turn to the language model below it.
A Reflex model cannot return an answer outside the set, and it cannot produce free text.
#What is available
| Model | What it reads | Status |
|---|---|---|
| revyn-turn | One turn of a sales or support call, with a little context | Preview |
Note
Preview means the model is trained and measured, and the numbers on its page are real. Weights and a hosted demo are not public yet, and it is not yet switched on inside Revyn Voice or phone agents. A human-annotated benchmark comes first.
#What Reflex is built from
Reflex models are fine-tuned from open encoder models. The base model, tokenizer, training loop and runtime are off the shelf. What Revyn builds is the question schema, which comes from the decisions our call engine already makes on every turn, and the training and evaluation data for it.
Something missing or wrong on this page? Tell us.