Kev 9B

A decision model, not a chat model: give it structured state and typed questions, get back calibrated choices, yes/no decisions and scores — no generated text, no output tokens to pay for.

Kev answers typed questions about structured state. Instead of generating prose it reads out probabilities — a choice among options you define, a yes/no decision, a score on a rubric — from a pointer head sitting on a Qwen3.5 base with a LoRA adapter. Each checkpoint ships with a fitted temperature, so the probabilities are calibrated out of the box.

That shape makes it the model for the decisions scattered through a production system: routing, classification, guardrails, evals. The questions are typed, so there is no parsing step and no prompt-injection surface in the output — and because decisions generate no tokens, output is free. Kev 9B is the mid-size of an Apache-2.0 family that also ships 0.8B, 4B and 27B checkpoints you can run yourself.

jaredpalmer/kev-9b

What’s it good for

Routing and classification

Choose a department, a priority, a pipeline branch — a typed choice with per-option probabilities, not text you have to parse.

Guardrails and checks

Calibrated yes/no decisions over structured state, with a fitted temperature so the probabilities mean what they say.

Scoring and evals

Rubric scores with honest uncertainty, at $0.03 per million input tokens and nothing for output. Cheap enough to score everything.

Benchmarks

Figures from the Kev-9B model card, October 2026. Decision accuracy is not comparable with generative benchmarks.

Accuracy, in-distribution87.3%decision-v7 locked test

Accuracy, out-of-domain85.2%transfer-v4 locked test

Calibration error0.034transfer-v4 locked test, at the fitted temperature

Confident errors1.4%wrong at ≥0.9 confidence, transfer-v4 locked test

Pricing & Specs

Latency is a choice, not a property of the model — pick the service tier your workload needs.

Input / 1M: $0.03
Cached input / 1M: $0.006
Output / 1M: $0.00
Context: 8.2K

Weights: Open — jaredpalmer/kev-9b, Apache-2.0
Released: September 2026
Base: Qwen3.5-9B-Base + LoRA + pointer head
Family: 0.8B · 4B · 9B · 27B
Precision on Gradiated: BF16
Output: Decisions only — no generated text

Getting Started

Kev speaks the System One API rather than chat completions: send state and named, typed questions through the TypeSafe SDK, pointed at our base URL.

Python — TypeSafe SDK

import os
from typesafe_sdk import Choice, TypeSafeClient

with TypeSafeClient(
    base_url="https://api.gradiated.com",
    api_key=os.environ["GRADIATED_API_KEY"],
    model="jaredpalmer/kev-9b",
) as client:
    response = client.system_one(
        state={"ticket": "I was charged twice. Please fix this today."},
        questions={
            "category": Choice(
                instructions="Choose the department for this ticket.",
                criteria={
                    "billing": "Charges or payments",
                    "technical": "Something is broken",
                    "other": "Anything else",
                },
            ),
        },
    )

print(response.choices["category"].choice)
print(response.choices["category"].probabilities)

Read the SystemOne SDK docs

Hardware

What Kev 9B runs on — our serving fleet first, then what the open weights support anywhere else.

NVIDIA RTX PRO 6000: What Gradiated serves on — one GPU per replica, BF16
Any 24 GB+ GPU: Self-host the 9B in BF16; the smaller family checkpoints run on less
Inference engines: Transformers with the Kev pointer head — LoRA on a Qwen3.5 base