Decision models: run Jev-style System One models on your own hardware

A decision model reads some text (the "state") and a set of typed questions, and returns a probability for every allowed answer in a single scoring pass. It never generates text. TypeSafe AI launched the category with its hosted model Jev on 15 September 2026. Since Ollama 0.35 (29 September 2026) you can run five decision models locally — nimble, tev1, tev1:0.8b, clef and clef-flash — through a new /v1/systemone endpoint. More than 70 open reproductions of Jev are now tracked on a community leaderboard, and the strongest of them score close to Jev on it.

This hub is the "can I run it and how" guide for the whole category.

Pick your starting point

What a decision model is good for

Think of it as a smart if-statement you can drop into ordinary code. Your program stays in control of the logic; the model only answers narrow questions about messy input:

• Route a support ticket to billing, technical or other.
• Gate an action: "does the customer explicitly ask for a refund?" → probability of yes.
• Score urgency, quality or risk on a rubric you define.
• Moderate or guard an LLM's input or output before it reaches a user.
• Choose a tool or a skill from a fixed list for an agent.

It is the wrong tool for anything that needs words out: summaries, drafts, explanations, chat. For those, you still want a regular local LLM — see the model library.

Which decision model can your hardware run?

Fit on your hardware: set your GPU or unified memory in the hardware analyzer and this page shows whether each model fits it. Check your hardware →

The short version, by published download size of the build you would actually run:

¹ The Index scores the Bespoke-Nimble-9B-v2 repository; which checkpoint the Ollama nimble tag packages is not stated in one place. ² Clef and Clef-Flash were released on 1 October 2026, after the 28 September snapshot, so they have no Index score yet. ³ Not on the 28 September snapshot; their authors' own scores are under Other benchmarks. ⁴ Liquid AI scored d1 with the official Index scorer itself (48.57 for d1-3B, 15.95 for d1-omni-600M). That is not a leaderboard result, so it is listed under Other benchmarks, not in this column. Decision Index snapshot: 28 September 2026. Scores are chance-corrected (0 = random guessing, 100 = perfect). Download size is a floor, not a verdict — use the hardware analyzer for a fit check that includes runtime overhead.

Tiny decision models (under 1 GB to 3B)

The models small enough for a phone runtime, a Jetson, a single-board computer or a laptop CPU. Each latency is the figure its source published, on the device and under the condition shown, so the rows are not a race: different devices, different request shapes, different people measuring. There is no ranking here.

The wider field: the top 25 downloadable models on the Decision Index

Index scores are chance-corrected (0 = random guessing, 100 = perfect); ECE is expected calibration error, lower is better. These are not the model authors' own benchmark numbers and are never plotted with them.

Other benchmarks (not comparable with the Index)

Scores the model authors published on their own suites. Each row is one model on one suite, with who ran it. A same-author benchmark is not independent even when its data is public, and rows on different suites measure different questions — never read this table as a ranking, and never set a number here against the Decision Index column above.

Benchmark built by the same team that built this model · Run by the authors on a public third-party harness · Authors' own internal benchmark — not independently reproduced. Every row is a different suite; the Decision Index column above is the only shared yardstick.

The honest caveats, up front

1. "Can't hallucinate" means "can't answer off-schema." The answer is always one of your options. It can still be the wrong option.
2. Ollama's confidence is not a correctness score. Ollama's own API reference defines it as how concentrated the probability distribution is, and says it is not calibrated correctness. Test thresholds on your own data before you trust them — here's how.
3. Two benchmarks, two stories. On Bespoke Labs' 13 human-labelled datasets, Nimble scores 75.7% against Jev's 76.0%. On the broader community Decision Index, Nimble v2 scores 39.6 against Jev's 57.9. Both are fair; they measure different things. We never put them on one chart.
4. Score is the weakest primitive. Nimble's public-benchmark accuracy is 81.6% on choice, 80.2% on yes/no and 54.6% on score questions.
5. Everything here is weeks old. Jev launched on 15 September, Ollama support on 29 September; Bosun v3.1 and DecisionBench on 22 September, OpenDecider on 27 September, Clef on 1 October, d1 on 7 October and THX-01 on 10 October. Expect model versions, limits and leaderboard positions to move. Every page in this cluster shows its last-verified date.

Keep going

• The model library — every decision model has its own page with hardware fit and sources
• Installing Ollama
• Check your hardware

Frequently asked questions

In this cluster

• Run Jev-style decision models locally with Ollama (Nimble, Tev1)Set up Nimble or Tev1 in Ollama 0.35 and send your first /v1/systemone request with curl, PowerShell, Python or JavaScript. macOS, Windows and Linux.→
• Open Jev reproductions beyond Ollama: Winnow, Decider, JevK5, AutoJev, LayaThe open decision models closest to Jev on the Decision Index, what hardware each needs, and how to serve them on a /v1/systemone endpoint your code already speaks.→
• Decision-model confidence, thresholds and calibration — use the probabilities safelyOllama's confidence score isn't a correctness score. How to measure calibration on your own data and set accept, escalate and abstain thresholds for local decision models.→
• Decision models on images: classify receipts, screenshots and forms locally with ClefSend photos and screenshots to Clef or Clef-Flash through Ollama's /v1/systemone: a working request, a Python helper for phone photos, and receipt triage.→
• How to evaluate a decision model on your own dataWhich evidence to trust, how to run DecisionBench on a pinned revision, and how to build a 200-row test from your own tickets to compare two models fairly.→
• Decision models on phones, Jetson and laptopsSmall decision models on a Mac, a Jetson, a small GPU or a CPU: what fits, how to serve each, and what the published latency does and does not tell you.→
• Ollama /v1/systemone errors and fixesWhat a 400, 404 or 413 from the decision endpoint means, and how to fix each.→
• Ollama /v1/systemone image errors: 400, 413, HEIC and sideways photosWhy an image sent to Clef or Clef-Flash is rejected or misread, and a script that finds the cause.→
• Decision models explained, with infographicsWhat Jev started, how the open reproductions compare, and the two benchmarks.→
• Decision request builderGenerates a valid curl, PowerShell, Python or JavaScript request and flags limit violations before you send it.→