[ Compare ]
Jev vs open decision models: Strands Decider and Cloudflare Clef compared
Published By JevHub
Two open-source rivals to Jev arrived in early October 2026: Strands Decider from AWS's Strands Labs, and Clef and Clef-flash from Cloudflare. Both can be run on your own hardware and use the same kinds of question as Jev. The speed and accuracy numbers so far come from the makers' own tests, run on different hardware, so they cannot be compared directly.
Side by side
| Feature | Jev (TypeSafe AI) | Strands Decider (Strands Labs) | Clef and Clef-flash (Cloudflare) |
|---|---|---|---|
| Released | 15 September 2026 | Reported 1 October 2026 (the project page is undated) | 1 October 2026 |
| Open or closed | Closed; used through TypeSafe's API | Open source, Apache 2.0; run it yourself | Open weights, Apache 2.0; or use it hosted on Cloudflare Workers AI |
| Size and base | Not published | 1.9 billion parameters on a Qwen3.5-2B base; Gemma 4 versions also released | Built on a 27 billion (Clef) and a 9 billion (Clef-flash) parameter base model, with Cloudflare's own training |
| Question types | Choice, Score, Noul (yes/no) | Choice, Noul, Score, the same three | Typed questions with choices from your schema; Cloudflare says it is compatible with the Jev API |
| Input | Text | Text; the Qwen3.5 version has optional image support | Text and images (Clef) |
| Speed (the maker's own test) | 70 to 500 ms (TypeSafe) | 115 ms median on one RTX 3090 graphics card; 153 ms on an M3 Pro laptop | Median 209 ms (Clef) and 38.8 ms (Clef-flash), against 524 ms for Jev in Cloudflare's table |
| Accuracy (the maker's own test) | Not stated here | 77.9% on the public JevBench set (180 of 231 tasks) | Leads on 7 of 10 benchmarks in Cloudflare's table (Clef 4, Clef-flash 3); Jev leads on 2 |
| Cost | $0.042 per million input tokens | Free software; you pay for the hardware you run it on | Open weights are free to run yourself; Workers AI pricing was not stated in the launch post |
| Request size | 64,000 tokens per request | Not stated | 64,000 tokens for Clef, according to Cloudflare |
Every figure in the Strands Decider and Clef columns comes from the makers' own pages and tests, on different hardware and tasks. Do not compare speeds across columns; use them to understand each model, then test on your own data.
The short answer
Jev is closed: you use it through TypeSafe's API and pay per token. Strands Decider and Clef are open: you can download them, run them on your own machines and tune them on your own data. That is the real difference. The cost of that control is that you run and look after the model yourself, unless you use Cloudflare's hosted version.
Nobody has yet published an independent, matched test of all of them against Jev. Treat the figures below as each maker's own claim.
Strands Decider
Strands Decider comes from Strands Labs, the experimental side of AWS's open-source Strands Agents project. The main model has about 1.9 billion parameters, uses the same three question types as Jev, and returns a confidence with every answer. You install it with pip install strands-decider, and it runs on NVIDIA graphics cards, Apple silicon or a plain CPU.
The project's page reports 77.9% accuracy on the public JevBench set (180 of 231 tasks) and a median response time of 115 ms on one RTX 3090 graphics card. It also lists larger Gemma 4 versions that score higher but need more memory. The page is candid about limits: it says differences smaller than about ten tasks are unresolved, and that its calibration claim covers only short classification tasks. It is a lab project, not a managed AWS service, so there is no service guarantee.
Cloudflare Clef and Clef-flash
Cloudflare released Clef, the larger and more accurate model, and Clef-flash, the faster one, on 1 October 2026. Both are open weights under Apache 2.0 and also run on Cloudflare's Workers AI platform. Clef can read images as well as text. Cloudflare also launched a service for tuning decision models on a customer's own data, run with Cloudflare's engineers at first.
In Cloudflare's own comparison table, Clef-flash has the lowest response time among the models that are as accurate (a median of 38.8 ms against Jev's 524 ms), and across ten benchmarks the two Clef models lead on seven and Jev on two. Cloudflare ran these tests itself, so independent confirmation is still to come.
Which should you pick?
Based on what the makers document:
- Choose Jev if you want to send a request and pay per token, with no machines to run.
- Choose Strands Decider if you want to run a small model on your own hardware, keep your data in house, or tune the model yourself.
- Choose Clef or Clef-flash if you already use Cloudflare, need images, or want an open model that is also available as a hosted service.
- Test before you commit. All of them take the same style of question, so you can run one labelled sample of your own data through each and compare.
Keep reading
More on how Jev compares and where it fits:
Common questions
- Is Strands Decider free?
- The software is free and open source under Apache 2.0. You pay for the computer you run it on.
- Is Cloudflare Clef really faster than Jev?
- In Cloudflare's own tests, yes: a median of 38.8 ms for Clef-flash and 209 ms for Clef, against 524 ms for Jev. Cloudflare ran those tests itself, so wait for independent results before relying on them.
- Can I run these models on my own computer?
- Yes. Strands Decider installs with pip and runs on a graphics card, an Apple silicon Mac or a CPU. Clef's weights are open on Hugging Face, and a hosted version runs on Cloudflare Workers AI.
- Can these replace Jev?
- They answer the same kinds of question, and Cloudflare says Clef is compatible with the Jev API. Whether they match Jev on your own data is something only a test on that data can show.
Sources
Written by JevHub from the sources below and checked against them on . Where a figure is a company's own claim, the article says so.
More comparisons: Jev vs ChatGPT, Claude and other LLMs.