Jev on Ollama: can you run it on your own computer?

No, not Jev. TypeSafe has not published Jev's weights, so there is no file for Ollama, or anyone else, to run on your own machine. Jev is used through TypeSafe's API.

What you can do is run a different model that answers the same kind of request. Ollama describes its decision models as "based on TypeSafe's Jev API": they take the same input and return the same shape of answer, so code written for Jev often needs only a new address.

• Is Jev open source?

Ollama's blog post, dated 29 September 2026, announced support for decision models in version 0.35. The new endpoint is POST /v1/systemone. You send some text as state and a set of named questions, and the model answers all of them in one request. Each question is one of three types, choice, noul (yes or no) or score, and each answer carries probabilities and a confidence value, as Jev's do.

Ollama names three models. nimble is a 9 billion parameter model made by Bespoke Labs, fine-tuned from Qwen3.5-9B and released under the Apache 2.0 licence. tev1 and tev1:0.8b are smaller models from Together AI, which Ollama marks as experimental. Ollama says more models are coming, including ones served from its own cloud.

These are the commands from Ollama's own blog post and documentation. First download the model:

ollama pull nimble

Then ask it a question. This example is Ollama's own:

curl http://localhost:11434/v1/systemone -H 'Content-Type: application/json' -d '{"model": "nimble", "state": "Our checkout has returned 500 errors since 9am.", "questions": {"label": {"type": "choice", "instructions": "Which label fits this ticket?", "criteria": {"billing": "Payments and refunds", "bug": "Software errors", "account": "Login and account access"}}}}'

In Ollama's documentation the answer picks bug with a probability of 0.9781 and a confidence of 0.8906. If you already use TypeSafe's Python SDK, Ollama's post says you point it at the local server with three settings:

export TYPESAFE_BASE_URL=http://localhost:11434

export TYPESAFE_API_KEY=ollama

export TYPESAFE_DEFAULT_MODEL=nimble

• How to use Jev

Close on the maker's own tests, not equal. The figures below come from Bespoke Labs, as shown on Ollama's nimble page, and compare nimble with Jev 1.13, not the newer Jev 1.14. Treat them as the vendor's claims.

• 13 public datasets. Mean accuracy of 75.7% for nimble and 76.0% for Jev 1.13.
• A held-out set. Nimble agreed with the reference labels 90.12% of the time, against 93.21% for Jev 1.13.0.
• A warning from the same page. "A probability of 0.9 doesn't mean the answer is right 90% of the time on your data." Ollama's documentation adds that confidence measures how concentrated the probabilities are, not how often the answer is correct.

• Jev benchmarks: every public test

Several limits matter before you swap a local model in for Jev:

• Short inputs. Nimble's model page says each prompt must fit in an 8,192-token context. Jev's default is 64,000 tokens per request.
• Text only by default. Images need a vision model such as Clef; nimble is a text model. Ollama's documentation does not mention a Box question type.
• No streaming, video, tools or generation controls. The answer is one JSON object.
• Local size limits. The text and question schema must fit in 64 KiB, and the whole request in 32 MiB.
• Cloud models. They need a paid plan or purchased credits.
• Your hardware. Ollama says nimble averaged 91 milliseconds per decision on an Apple M5 Max laptop. That is Ollama's own number from one demo, and speed will differ on your machine.

Based on what is published today:

• Use Ollama and nimble if the text must stay on your own machine, you want no per-call charge, or you want to build and test offline. You provide the computer, and you should check accuracy on your own data.
• Use Jev if you want the best published accuracy, long inputs, image experiments or no hardware to run. You pay TypeSafe per token and your text goes to its servers.
• Try both on the same sample. Because the request format matches, you can run one labelled sample through each and compare accuracy, speed and cost before you choose.

• Jev vs open models: can you skip Jev?
• Jev pricing

This matters because it makes Jev's request format a shared standard rather than one company's API. Your code stays the same whether the answer comes from TypeSafe or from your own laptop. It also means Jev now has a free local stand-in for development and testing, which lowers the cost of trying the whole approach. Whether nimble is accurate enough to replace Jev in production is something only a test on your own data can show.