Open Jev reproductions beyond Ollama

Ollama ships five decision models. The community has published about 70 open reproductions of Jev, and several score far higher on the Decision Index — Winnow-12B (50.0) fits a 16 GB card at Q8_0, and AutoJev-27B (56.4) comes within 1.5 points of Jev (57.9) but needs about 49 GiB for its BF16 weights. Almost all of them expose the same POST /v1/systemone request shape, so code you wrote against Ollama keeps working — you change the URL and port.

When to leave Ollama's decision models

Stay with nimble / tev1 if they're accurate enough on 50–100 of your own labelled examples. Move on when:

• your questions are harder than routing — multi-step judgments, knowledge questions, long policies — where the Index gap between Nimble (39.6) and the leaders (50–57) shows up;
• you need more than 26 options in a choice (Ollama's limit; Decider and Jev go to 255);
• you need non-English input (Laya's multilingual router covers 100+ languages; most others are English-first);
• you need images in the state (Ollama's clef and clef-flash accept images, and so do AutoJev, Intern-Decision and Winnow; Ollama's endpoint does not take them for nimble or tev1). See Decision models on images.

The shortlist

Only models with open weights, a permissive licence, and a documented local runtime made the cut. Decision Index 0.2.1 snapshot of 28 September 2026 (chance-corrected; 0 = random, 100 = perfect).

² Laya is an encoder built for fast routing and guardrails in many languages, not a general reasoner — its Index score reflects that. It's on the list because nothing else covers 100+ languages locally.

Left off on purpose: the top-ranked Surogate Rune 26B-A4B (57.4) because its Hugging Face repo listed no GGUF file when we checked on 30 September; the JPT family because it is CC-BY-NC-4.0 (non-commercial); and every "inference technique" entry on the Index, because those are decoding methods applied to stock models rather than weights you download. We'll add Rune once its files are published.

Added in October 2026. Five more families now have pages, and Rune is among them: its weights are out, though the repo holds BF16 safetensors rather than the GGUF its name suggests, and it is gated. This table renders from each model's record, so it stays in step with the model pages. One licence note: Liquid AI's d1 models ship under the LFM Open License, which allows commercial use only below a company-revenue threshold, so they are not permissive in the sense the table above uses.

Rendered from each model's record. "Index" is the community Decision Index score where the model is on the board; models scored only by their authors show "—" here, and their own benchmarks are on their pages.

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 →

Pattern 1 — GGUF + llama.cpp-based server (Winnow)

The most hardware-flexible path. Winnow's own server is built on llama.cpp and exposes /v1/systemone and a normal /v1/chat/completions from the same loaded model — one download gives you a decision model and a chat/vision model.

Tested setups, per its README: Linux with an NVIDIA GPU (RTX 5070 Ti, CUDA 12.8+), Apple Silicon with 24 GB or more of unified memory, and an NVIDIA Docker container. The Q8_0 build was run with a 64K context on a 16 GB RTX 5070 Ti.

# bash — Linux (NVIDIA) or macOS (Apple Silicon)
git clone https://github.com/EldanRing/winnow-inference
cd winnow-inference
python3 scripts/setup.py                             # downloads the model (~12.9 GB)
python3 scripts/serve.py --profile 5070ti-64k        # NVIDIA
# python3 scripts/serve.py --profile apple-silicon   # Mac

It listens on http://127.0.0.1:8091. Your Ollama request works unchanged apart from the URL:

curl http://127.0.0.1:8091/v1/systemone \
  -H 'Content-Type: application/json' \
  -d '{"state":"The store accepts returns within 30 days. This item was bought 12 days ago.",
       "questions":{"eligible":{"type":"noul","instructions":"Is the item eligible for return?"}}}'

Check the repo's quickstart for whether model is required and for the current profile names — this project is two weeks old and moves quickly.

Pattern 2 — Python package with a built-in server (Decider, JevK5, AutoJev)

These ship their own inference code and a small HTTP server that speaks TypeSafe's format. You need a working PyTorch install with GPU support first.

Decider — the most downloaded open decision model (over 255,000 downloads of the 2B in its first two weeks). Four sizes share one interface.

# Python 3.11+; needs torch, transformers>=5 and flash-linear-attention
# (it runs without flash-linear-attention, but several times slower)
from decider.infer import Decider   # the decider/ package ships inside the model repo

d = Decider("Mapika/decider-2b")
d.decide(
    "My card was charged twice for the same purchase.",
    [{"question": "Which department should handle this?", "options": ["billing", "technical support", "sales"]},
     {"question": "Does this need a refund action?", "options": ["no", "yes"]}],
)

For an HTTP endpoint, run decider.serve with DECIDER_MODEL pointing at the downloaded model folder; the official typesafe-sdk works against it unchanged with TYPESAFE_BASE_URL set to that server. Decider allows 2–255 options per question, and abstain_below=t makes it return None instead of an answer when confidence is under t.

JevK5 — one of the best-calibrated 4B models on the Index (expected calibration error 0.027). English only; inputs over 16,384 tokens are refused rather than cut.

# bash — needs a CUDA GPU with ~9 GB free for BF16
jevk5-serve --model alibiserikbay/JevK5 --port 8090

No NVIDIA card? The alibiserikbay/JevK5-GGUF builds run under llama.cpp on NVIDIA, AMD, Intel and Apple GPUs or the CPU. Per its card, the Q8_0 file matches the full-precision answer on 229 of 231 public JevBench items, and Q4_K_M on 221.

AutoJev-27B — the strongest open model you can actually download today. Its card asks for a GPU with room for about 49 GiB of BF16 weights plus runtime overhead — in practice a 64 GB-class or larger workstation card.

# bash — Python 3.12+ and uv
git clone https://github.com/denis-pplx/autojev.git
cd autojev
uv sync --frozen --python 3.12
uv run hf download denis-pplx/autojev-27b --local-dir checkpoints/selected
AUTOJEV_CHECKPOINT=checkpoints/selected uv run autojev-serve

It serves a browser playground at http://localhost:8000 and POST /v1/systemone with optional base64 images. Set AUTOJEV_API_KEY to turn on authentication — do that before exposing it to a network.

Pattern 3 — A tiny encoder for routing in any language (Laya)

pip install laya            # extras: laya[serve], laya[mcp], laya[langchain], laya[onnx]

from laya import Router

router = Router()   # downloads a checkpoint on first use
result = router.predict(
    "Hi, we were billed twice for March. Please refund the duplicate today or we will cancel our plan.",
    {"department": {"type": "choice", "instructions": "Which department should handle this?",
                    "criteria": {"billing": "invoices, payments, refunds",
                                 "technical": "bugs, outages, system errors",
                                 "other": "everything else"}}},
)
print(result["answers"]["department"]["choice"], result["routing"]["model"])

The router detects the script and language and sends non-English text to the multilingual checkpoint automatically. It's the fastest thing on this page and the least capable at reasoning — use it as a first-pass router, not a judge.

/v1/systemone-compatible servers

You do not have to run a decision model in the server its authors ship. Several general-purpose servers now speak the same request shape: a state, plus questions keyed by name, each a choice, noul or score. Ollama has served it since v0.35; the llama.cpp server, LM Studio and Hugging Face's transformers serve document it too. The model-specific servers (Bosun's jev-compatible-server, opendecider serve, THX-01's own server) remain the only route for features the general servers do not cover, such as THX-01's extra question types.

General-purpose servers first, read from each engine's documentation; then the servers that ship with one model family, read from that model's card and listed on its page. "—" means the documentation does not say. Max options is what the server or card states; a model can accept fewer than its server does.

How to choose:

• Already on Ollama or LM Studio? Stay there for the models they load. Ollama's endpoint is the reference most guides on this site assume, and LM Studio adds a second, OpenAI-style /v1/decisions route for clients built on the OpenAI SDK.
• Need images or audio? The llama.cpp server reads images for any decision model that ships a multimodal projector, and its documentation names one architecture (lfm2-d1-omni) for audio. Ollama reads images only through Clef and Clef-Flash.
• Need many options per question? Ollama caps a choice at 26 options. The llama.cpp server documents per-model limits that reach 255 for some of the models it names. Check the limit on the model's page before you design a 40-way router.
• Evaluating a model nobody has packaged yet? transformers serve scores the allowed answers of any chat model with a chat template, in one batched forward pass. That makes it a handy harness, but a general chat model scored this way is not a trained decision model: measure its calibration before you trust its probabilities (see the confidence guide).
• Want THX-01's number, excerpt or citation types? Only its own server returns them.

Two request shapes are in use. /v1/systemone takes a state and an object of named questions. The OpenAI-style /v1/decisions that LM Studio documents (and Rune's server uses) takes an input and a list of questions. They answer the same kind of question, but a client written for one will not talk to the other without changes. Rune's page says which one it expects.

The request builder writes valid /v1/systemone requests for the models it supports.

Sources, checked 2026-10-14: Ollama System One API (v0.35.0+, choice up to 26 options) · llama.cpp server: /v1/systemone, per-model option limits, image and audio input · LM Studio: /v1/systemone · LM Studio: OpenAI-style /v1/decisions · transformers serve: /v1/systemone for any chat model.

Which one for your hardware?

This table is guidance from the published file sizes and each project's tested setups. Your analyzer result is the verdict for your exact machine. If the model you need doesn't fit what you have, renting a GPU by the hour is often cheaper than buying for an occasional batch job — compare in the cost calculator.

For phones, Jetson boards and plain CPUs, the smaller models added in October 2026 change the picture; decision models on edge hardware goes device by device. Before you commit to any of these, test your top two on your own data, as the evaluation guide describes.

Train your own

Two of these projects publish their full recipe:

• Together AI's Tev1 — data recipe and training code at github.com/togethercomputer/tev1, with a Together blog post on training your own classifier for about $17.
• Decider — code, data registry and training scripts at github.com/Mapika/decider.

A decision model fine-tuned on a few thousand of your own labelled decisions usually beats a bigger general one on your task. Our fine-tuning guides cover the basics of LoRA on a single GPU.

Frequently asked questions

Verified against: Decision Index 0.2.1 data (generated 28 Sep 2026); Hugging Face model cards and GitHub READMEs read 30 Sep 2026.

Keep going

• Decision models (Jev-style System One): run them locally→
• Run Jev-style decision models locally with Ollama (Nimble, Tev1)→
• Decision-model confidence, thresholds and calibration — use the probabilities safely→
• Decision models on images: classify receipts, screenshots and forms locally with Clef→
• How to evaluate a decision model on your own data→
• Decision models on phones, Jetson and laptops→
• Ollama /v1/systemone errors and fixes→
• Ollama /v1/systemone image errors→
• Decision request builder→