Skip to content

Skena Labs · experiment

Meet Kvick. Decisions in 80 ms.

A small, fast model that answers yes or no, multiple choice and 1 to 5 ratings about any text or image, each with a calibrated probability. It speaks the Jev API, so code written for Jev works by changing one address.

Request accessSee the comparison

experimental · free to try · request access

A silent reel from the Kvick film: Kvick arrives beside Servant, reads a café receipt and answers four questions about it, answers a stream of support emails, and the benchmark bars grow to 82.4% against Jev 1.13's 76.0%.
on a public, human-labelled benchmark
82.4%
on a public, human-labelled benchmark
Jev 1.13: 76.0%
for one decision, end to end
75 ms
for one decision, end to end
network included
on 1,800 questions about real images
93.6%
on 1,800 questions about real images
six public image benchmarks
datasets ahead of Jev
13 of 13
datasets ahead of Jev
two of them by a few records

what it does

Three kinds of answer, each with a probability.

Kvick never writes text. It reads your question once and answers with the probability of every possible answer, so you set the thresholds.

  • Yes or no

    noul

    “Is the customer asking for money back?”

    The probability of yes.

  • Multiple choice

    choice

    “Which team should handle this: billing, technical or sales?”

    The answer, a probability for every option, and a confidence.

  • A 1 to 5 rating

    score

    “How frustrated is the customer?”

    The expected rating, a probability per level, and a confidence.

POST /v1/systemone
Authorization: Bearer <your key>

{
  "model": "kvick-1",
  "state": "My card was charged twice, please refund one.",
  "questions": {
    "refund": {
      "type": "noul",
      "instructions": "Is the customer asking for money back?"
    },
    "department": {
      "type": "choice",
      "criteria": { "billing": null, "technical": null, "sales": null }
    },
    "frustration": {
      "type": "score",
      "instructions": "How frustrated is the customer?",
      "criteria": ["calm", "mildly annoyed", "annoyed", "frustrated", "furious"]
    }
  }
}
The Jev API, unchanged: a TypeSafe SDK or your own Jev client works by pointing it at Kvick's address with your key.

against jev, on a public benchmark

More accurate. Fast enough. And it sees.

Accuracy · 13 human-labelled datasets
Accuracy · 13 human-labelled datasetsKvickJev 1.13
Overall, 13 datasets82.4%76.0%
Datasets Kvick never trained on76.5%70.8%
Yes or no questions91.0%84.6%
Multiple choice86.6%82.9%
1 to 5 ratings60.9%50.1%
Datasets ahead13 of 130 of 13
Speed, end to end
Speed, end to endKvickJev
One decision75 ms~100 ms
8 choices in one call109 ms114 ms
14 yes or no in one call145 ms111 ms
13 questions on a long document (13k tokens)~570 ms~270 ms

It sees images too.

Send an image as the state and ask the same three kinds of question. On 1,800 questions drawn from six public image benchmarks (POPE, TextVQA, A-OKVQA, AI2D, MMBench and ScienceQA), Kvick answers 93.6% correctly. An image and four questions take about 140 ms.

How these numbers were made, and where they stop: the method.

method and limits

How we measured.

  1. 01The benchmark is Bespoke Labs' public suite (github.com/bespokelabsai/nimble, docs/PUBLIC_BENCHMARKS.md): 13 human-labelled datasets, 3,880 records. We rebuilt every dataset byte for byte from the committed record ids and scored with the suite's own definitions.
  2. 02Jev's figures are the ones Bespoke Labs published for Jev 1.13. We used no Jev output at all.
  3. 03Nothing was fitted on the suite. Calibration comes from our own split, training rows that share text with a suite record were dropped, and five of the datasets were never trained on by any version.
  4. 04Two leads are thin: BoolQ by about two records and VitaminC by about ten. Run-to-run noise is about 0.3 points.
  5. 05Kvick's speed is wall time from Stockholm to one NVIDIA RTX PRO 6000 in europe-west4, about 45 ms of network included. Jev's is from TypeSafe's published cookbooks. On long documents Jev is still about twice as fast.
  6. 06Kvick is a fine-tune of an open-weight Qwen mixture-of-experts model with 3B active parameters: an average of three fine-tuned checkpoints.

Jev and TypeSafe are trademarks of their owners. Skena Labs is not affiliated with TypeSafe or Bespoke Labs.

try kvick

Request access. Get your key.

  1. 01

    Create a free Skena account.

  2. 02

    Request Kvick on your account's Early access page.

  3. 03

    Once you're in, your key and the address to send requests to appear right there.

Request access
  • An experiment, not a product: no uptime guarantee, and it may change or stop.
  • The image model starts on demand, so the first image request after a quiet spell can take a few minutes.
  • Please do not send personal or sensitive data while it is experimental.