Skip to content

Published by DecisionNode (Bynn Intelligence, Inc.). Not affiliated with, endorsed by or operated by TypeSafe AI. About this site

Jev, compared

Jev alternatives

On measured accuracy, the strongest alternative is DecisionNode MoE: 0.816 against 0.803 for Jev on 15 text suites. Jev still leads on knowledge questions. Eight models, ranked, with the method shown.

Facts checked

Why teams look past Jev

Jev does one job well: typed answers about text. Teams start looking elsewhere when the job grows past what TypeSafe documents.

  • Inputs that are not text. Jev reads text and JSON only, so photos, screenshots and video need a separate model first.
  • Answers beyond three types. Jev returns choices, scores and yes/no; counts, positions and boxes are out of scope, and TypeSafe advises doing numbers in code.
  • Adversarial input. TypeSafe lists text written to steer the model as a known weak spot, which matters for agents and user content.
  • Volume pricing. Jev's documented API has no batch price and no free tier (Jev pricing).

What to check in an alternative

  1. What it reads: text only, or images and video too.
  2. The answer types your code needs.
  3. Accuracy on task families like yours, not only the mean.
  4. How it holds up when the input tries to steer the answer.
  5. Price per decision at your request size, and batch pricing.
  6. Where it runs: a hosted API, or open weights you serve yourself.
  7. What happens to your data: training and retention.

The ranking

Mean accuracy on 15 text suites, with prompt injection and Photos beside it. Jev is listed as the reference.

Decision models ranked by mean accuracy on 15 text suites
System 15-suite mean Prompt injection Photos
DecisionNode MoE 0.816 0.764 0.918
TypeSafe Jev 1.13 reference 0.803 0.600 text only
Cloudflare Clef 27B 0.786 0.582 0.882
DecisionNode Flash 0.784 0.747 0.900
Cloudflare Clef-flash 9B 0.759 not published 0.905
pplx-decider-27b 0.746 0.318 0.891
Cygnet (Gemma 4 12B) 0.717 0.327 not published
Kev-4B 0.717 0.652 not published
Strands Decider 2B 0.622 not published not published

Preliminary results, October 2026. 15 text suites, 25,134 items, development splits. 95% intervals for the DecisionNode models and Jev come from a bootstrap that resamples whole inputs. Jev (jev-1.13.0) was called through its official API; open models ran as their authors serve them. DecisionNode models are release candidate 1.

“Leads” and “trails” are used only where the two 95% intervals do not overlap. Source: decisionnode.com/benchmarks. DecisionNode is not affiliated with TypeSafe AI, Cloudflare or the other model publishers. Their names identify the systems measured.

Each alternative

  1. 1

    DecisionNode MoE

    Bynn Intelligence

    The highest 15-suite mean measured (0.816) and the top prompt-injection score among the systems with a published result (0.764). Reads text, JSON, images and video frames, and answers choice, score, yes/no, number, point and box questions. $0.042 per 1M input tokens, the same as Jev.

    Best for: Text and image decisions where accuracy and adversarial input matter, such as agent guardrails and moderation.

    Jev vs DecisionNode, in full
  2. 2

    Cloudflare Clef 27B

    Cloudflare

    0.786 on the 15 text suites, level with DecisionNode Flash, and 0.882 on Photos. On Rendered scenes it scores 0.783, above both DecisionNode models; on prompt injection it scores 0.582.

    Best for: Questions about rendered or synthetic scenes.

  3. 3

    DecisionNode Flash

    Bynn Intelligence

    0.784 on the 15 text suites and 0.900 on Photos, at $0.035 per 1M input tokens, or $0.0175 in batch. DecisionNode publishes 19 ms end to end for a short Flash decision, against 1,300 ms for GPT-5.6 Luna.

    Best for: High-volume decisions where price per call matters more than the last point of accuracy.

    Flash pricing
  4. 4

    Cloudflare Clef-flash 9B

    Cloudflare

    Only its 15-suite mean (0.759) and its image results are published: 0.905 on Photos and 0.742 on Rendered scenes.

  5. 5

    pplx-decider-27b

    Open model

    An open model. A 0.746 mean, with the top scores measured on Held-out decisions (0.959) and Rendered scenes (0.809), but the lowest prompt-injection score (0.318) and 0.492 on Long inputs.

    Best for: Self-hosting, when inputs are short and come from trusted sources.

  6. 6

    Cygnet (Gemma 4 12B)

    Open model

    An open model built on Gemma 4 12B. A 0.717 mean, 0.913 on Short classifications and 0.327 on prompt injection.

    Best for: Self-hosted classification of short, trusted inputs.

  7. 7

    Kev-4B

    Open model

    An open model. A 0.717 mean, and 0.652 on prompt injection, the third-best score after the two DecisionNode models.

    Best for: A small model to self-host when inputs may be adversarial.

  8. 8

    Strands Decider 2B

    Open model

    An open model with a 0.622 mean, the lowest measured. Its published training data contains items of two of the test sets, so even that figure may be inflated.

General chat LLMs can also be prompted to return a decision as JSON. They are not ranked here, because no accuracy figure is published for them on these suites.

Which one, by use case

You needLook atWhy
Images or video DecisionNode MoE or Flash Jev reads text only; MoE scores 0.918 on Photos.
Agent guardrails, untrusted input DecisionNode MoE 0.764 on prompt injection, against 0.600 for Jev.
High volume at the lowest price DecisionNode Flash, batch $0.0175 per 1M input tokens.
Questions that need world knowledge Jev 0.820 on MMLU-Pro, the best measured.
Open weights you serve yourself Kev-4B or pplx-decider-27b Kev holds up better on adversarial input; pplx scores higher on short, trusted tasks.

Questions about Jev alternatives

What is the best alternative to Jev?

On measured accuracy, DecisionNode MoE: 0.816 as the mean of 15 text suites against 0.803 for Jev, and the top prompt-injection score measured. It also reads images and video, which Jev cannot. Jev still leads on knowledge questions, so test on your own task family before you switch.

Is there a free alternative to Jev?

DecisionNode has a free tier: 50 DecisionNode Flash requests, 20 DecisionNode MoE requests and 10 booster calls a day, on the first workspace. Open models such as Kev-4B and pplx-decider-27b run on your own GPUs: no fee per token, but you pay for the hardware.

Is there a Jev alternative that reads images?

Yes. DecisionNode MoE and Flash, Cloudflare Clef 27B, Cloudflare Clef-flash 9B and pplx-decider-27b all have results on the Photos suite. Jev reads text only.

Can a general LLM replace Jev?

A chat LLM can be prompted to return JSON, but it generates its answer as text, which is slower and can break the requested shape. DecisionNode publishes one timing: GPT-5.6 Luna took 1,300 ms on a short decision that DecisionNode Flash answered in 19 ms, timed from the same client. No accuracy figure is published for chat LLMs on these suites.

Run your own decisions on DecisionNode

Send text, JSON or images with typed questions and get probabilities back. Free tier: 50 Flash and 20 MoE requests a day.

Try DecisionNode