Jev vs the open-weight decision models: how they compare

Decision Models Digest, Oct. 28, 2026

Jev, from TypeSafe AI, is a closed model. You can only use it through TypeSafe's API or a reseller such as Vercel's AI Gateway. Open-weight decision models publish their weights, so teams can download them and run them on their own hardware. That matters for companies that can't send data to an outside service, or that make enough decisions for self-hosting to cost less.

Which open-weight decision models matter most?

Three large companies have released open-weight models, all under the Apache 2.0 license:

Perplexity's pplx-decider, a 27-billion-parameter model with a smaller flash version.

Cloudflare's Clef and Clef-flash, at 27 billion and 9 billion parameters. Cloudflare hasn't published the training data.

AWS's Strands Decider 2B, a small model released with its weights, training data and training scripts.

Independent developers have built many more. The strongest on the benchmarks include Imajev, JevK5, Plumb and Mapika's decider-4b. Jared Palmer's Kev copies Jev's design most closely. Nace.AI's Drex 1.5 is also open, though its newer Drex models are hosted only.

How do they score against Jev?

It depends on the benchmark.

On JevBench version 1.4.2.2, three small open models ranked above Jev. Imajev-4B scored 67.37, Plumb-4B 65.84 and decider-4b v2 64.13, against Jev's 63.29. But JevBench weighs speed and cost as heavily as accuracy. On its accuracy measure alone, Jev beat decider-4b v2, 53.1 to 49.4.

On DecideBench, which measures accuracy on 400 decisions, Jev scored 98.0%. The best self-hosted model, imajev-4b, scored 95.0%. Clef scored 94.0% and Kev-4B 78.2%.

On the Artificial Analysis Decision Index, pplx-decider scored 56.2 and Jev 57.5. The leader there is Liquid AI's d1-ultra at 65.2, but it's a closed, hosted model.

The short version: the best open models are close to Jev, and a few beat it on combined scores. On raw accuracy, Jev still holds a small lead over the open models.

Can I switch from Jev without rewriting my code?

For some models, yes. Kev serves the same API as Jev, so TypeSafe's own software kit works with it after changing one address. Model listings say Nace.AI's Drex accepts Jev's request format, and DecideBench reports that Clef does too. OpenAI's Decisions API uses its own format, so switching to it takes more work.

What does it take to run one yourself?

Hardware needs vary with size. The Register reports that Clef needs about 85 GB of GPU memory and Clef-flash about 41 GB. Small models need far less. Imajev-4B was benchmarked on a single consumer graphics card, and AWS reports Strands Decider 2B answering in under 100 milliseconds on an RTX 3090.

There are no per-token fees, but the hardware isn't free. DecideBench estimated imajev-4b at about $29 per million tasks on rented GPUs, against $32 for Jev through a hosted gateway.

Which should I choose?

Jev and other hosted models are the simplest option. There's nothing to run, and the top hosted models remain the most accurate. Open-weight models make sense when data can't leave your systems, when you make decisions in very large volumes or when you want to fine-tune a model for your own task. Test any candidate on examples from your own work before choosing. The benchmarks measure general tasks, not yours.