On 1 October 2026 Amazon Web Services released Strands Decider 2B and Cloudflare released Clef and Clef-flash: decision models with open weights that take a situation and a list of options and return the probability of each. Jev from TypeSafe, used through an API, had come out on 15 September, and VentureBeat described CLM-8B from Stanford and Nvidia on 25 September. If a flow of yours uses a language model to sort requests or filter a chatbot's answers, Decider, the smallest, can be tried on your own data on an ordinary computer: before deciding anything else, measure how often it gets your cases wrong.
How a decision model works
A language model writes its answer a piece at a time; a decision model makes a single pass over the text and gives a score to each option you pass it, in the format the code expects, with the probability alongside. Classifiers have existed for some time: according to Cloudflare, what is new is that these models work on any set of inputs without being retrained for every new category. According to TechCrunch, dozens of similar models have appeared since Jev's debut, and in the week of Decider OpenAI also announced a comparable offering.
The models side by side
| Model | Base | Weights licence | Where it runs |
|---|---|---|---|
| Strands Decider 2B (Amazon) | Qwen3.5-2B | Apache 2.0 | Locally: Nvidia GPU, Mac or CPU |
| CLM-8B (Stanford and Nvidia) | Qwen3-8B | Apache 2.0 | Weights and code to download |
| Clef and Clef-flash (Cloudflare) | Qwen3.8-27B and Qwen3.5-9B | Apache 2.0 | Workers AI or downloadable weights |
| Jev (TypeSafe AI) | Not disclosed | Not stated | TypeSafe API, early access |
Every figure is claimed by whoever publishes the model, on different test sets and machines, and cannot be ranked against the others. TypeSafe quotes 70 to 500 ms for Jev, from tests run on laptops on the US West Coast, while in Cloudflare's table its median is 524.1 ms, against 209.3 for Clef and 38.8 for Clef-flash. The authors of CLM-8B claim up to 9 times Jev's speed, with 95.2% against Jev's 99.2% on BFCL v4, a tool-calling test.
Cost, privacy and hardware
Open models carry no licence fee, and the sources give no hardware costs. Among the services, Jev costs $0.042 per million input tokens with free output (TypeSafe does not rule out that the price is subsidised); on Workers AI Clef-flash costs $0.09 and Clef $0.24, and Cloudflare lists no output price. With 10,000 requests a day of 500 input tokens each, that is 5 million tokens: at list price about $0.21 a day with Jev, $0.45 with Clef-flash and $1.20 with Clef [illustrative estimate, to be confirmed against your own project]. At these volumes the spend is small, and in our view the reason to keep the model in house is privacy.
With the weights in house the text is processed on your own computer; with an outside service the data passes through a vendor: Cloudflare writes that it neither reads nor stores requests and responses, nor trains on them, except for fine-tuning chosen by the customer, and TypeSafe's service is currently based in the United States.
Decider is the smallest: Amazon quotes a median of about 115 ms on ordinary hardware, an RTX 3090 in the blog's chart, and about 153 ms on an M3 MacBook for small tasks. Clef's weights are about 55 GB and its card cites a single Nvidia H200 as its only test. At 115 ms per decision, one at a time, 24 hours give about 750,000 decisions, a theoretical ceiling for short texts because the time grows with length [illustrative estimate, to be confirmed against your own project].
How to set up a test
- Pick a decision with few possible answers, for example routing a request to one of three or four teams.
- Collect at least 200 real cases already labelled by a person; if they contain personal data, they stay on the test computer. With 200 cases the uncertainty on accuracy reaches about 7 percentage points in the worst case [illustrative estimate, to be confirmed against your own project]: little to tell two close models apart.
- Install Decider with
pip install strands-decideron a laptop or a PC with an Nvidia card, run the cases, note the time per decision on your own machine and compare where it goes wrong with what you do today and with the language model you already use. - Set the probability threshold on your cases and pass decisions below it to a person. Do not expose Decider's HTTP server to the network: according to the card it has no authentication and is meant for local experiments.
Where we start
In AI automation and chatbots we integrate language models into existing workflows, with control over costs and answer quality. Sorting a ticket or filtering an assistant's answer is a step where a decision model can sit beside, or replace, a call to a model that writes: for a client we would start from their own labelled cases and compare costs and errors with the tool they use today.
In mobile apps the open point is the phone. No source documents these models on a device, and for phones Android describes a different model, Gemini Nano, that produces text and does not mention classification. To bring a PyTorch model into a Flutter app we have Fluttorch, our open-source library: the export writes the model, its manifest and the reference outputs together, and the parity gate fails the build if the numbers move further than the export recipe allows, for example after quantisation. A probability threshold chosen on your data holds only while the probabilities stay what they were. Fluttorch does not train or optimise models, and we do not promise that any of these will export to a phone: that would be the first check.
What we do not know yet
- How well they handle Italian: the sources report no tests, and Brooker, one of Decider's authors, cites to TechCrunch understanding of languages among the things not to degrade.
- What hardware Clef-flash and CLM-8B need, and how fast Decider is on a CPU alone.
If you want us to look at a sorting or checking flow of yours, start from the page on AI automation and chatbots or get in touch.
Sources
- TechCrunch: Amazon releases its own Jev clone as decision models flood the web (1 October 2026)
- Strands Agents: Introducing Strands Decider 2B: a small, open source, decision model (1 October 2026)
- Hugging Face: StrandsAgents/strands-decider-2B-hobson-v19 (model card)
- VentureBeat: Stanford and Nvidia's open CLM-8B caches reusable agent actions and runs up to 9x faster than Jev in tests (25 September 2026)
- Cloudflare Blog: Introducing Clef: our open-source decision models, and new RL fine-tuning platform (1 October 2026)
- Hugging Face: Cloudflare/clef (model card)
- Cloudflare Workers AI docs: Pricing (1 October 2026)
- TypeSafe AI: Introducing System One Models & Jev (15 September 2026)
- Android Developers: Gemini Nano (8 September 2026)
