Decision Models · Run locally

Decision models, close to your data.

Choose an open-weight model, check the hardware it needs, and serve it through a Jev-compatible endpoint on a machine you control.

01

Your data stays on your machine

Prompts and decisions stay within the environment you choose.

02

No network round trip to the model

Measure response time on the device and network you plan to use.

03

One endpoint shape across local runtimes

Call the same typed /v1/systemone shape used by hosted systems.

04

No telemetry; loopback by default

The local service binds to 127.0.0.1 unless you change it.

Published benchmark

JevBench top 10 (text)

Capability rank and score are read from the current catalogue. Model licences and local hardware requirements still apply.

Published benchmark

ImageJevBench top 10 (image)

Capability rank and score are read from the current catalogue. Model licences and local hardware requirements still apply.

Model not offered

This model is not offered in the installer while its licensing review is pending.

Local model catalogue

Choose a local model

18 models have an executable install recipe. Others are marked setup in preparation. Model pages show runtime, memory, and licence details.

Recipe ready · unverifiedCommercial usetextimage

AutoJev-27B

27B text and image model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Recipe ready · unverifiedCommercial usetext

Bobcat Flash 1.2

26B text model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Recipe ready · unverifiedCommercial usetext

decider-12b v2

12B text model.

Hardware estimate: Apple Mac mini M5 Pro (18-core CPU, 20-core GPU, 64 GB, 1 TB) · unverified
Recipe ready · unverifiedCommercial usetext

decisio v0.8.0 on gemma-4-31B-it

31B text model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Recipe ready · unverifiedCommercial usetext

deck-31B

31B text model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Install testedCommercial usetext

H2O-Lightning-4B v1.1

4B text model.

Hardware estimate: NVIDIA GeForce RTX 5090 (32 GB) workstation · install tested
Recipe ready · unverifiedCommercial usetextimage

Imajev-4B

4B text and image model.

Hardware estimate: Apple Mac mini M6 (16 GB, 256 GB) · unverified
Recipe ready · unverifiedCommercial usetextimage

Jev-Omni

12B text and image model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Recipe ready · unverifiedConditional termstextimage

Jevify Gemma 4 26B-A4B

25.2B text and image model.

Hardware estimate: Apple Mac Studio M5 Ultra (30-core CPU, 64-core GPU, 96 GB, 1 TB) · unverified
Non-commercial onlyRecipe ready · unverifiedNon-commercialtextimage

JPT-4B

4B text and image model.

Hardware estimate: NVIDIA GeForce RTX 5090 (32 GB) workstation · unverified
Recipe ready · unverifiedCommercial usetextimage

NeoHorse Jev 4B

4B text and image model.

Hardware estimate: Apple Mac mini M5 Pro (15-core CPU, 16-core GPU, 24 GB, 512 GB) · unverified
Recipe ready · unverifiedCommercial usetext

Quyet-1.0-Large

31B text model.

Hardware estimate: Apple Mac Studio M5 Ultra (30-core CPU, 64-core GPU, 96 GB, 1 TB) · unverified
Recipe ready · unverifiedCommercial usetextimage

reflex 4B

4B text and image model.

Hardware estimate: Apple Mac mini M5 Pro (18-core CPU, 20-core GPU, 64 GB, 1 TB) · unverified
Recipe ready · unverifiedConditional termstext

René-1 31B FP8

31B text model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Recipe ready · unverifiedCommercial usetextimage

shisa-de-1

25.2B text and image model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Recipe ready · unverifiedCommercial usetextimage

Surogate Rune 26B-A4B v3

26.5B text and image model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Non-commercial onlyRecipe ready · unverifiedNon-commercialtext

torchcast-decision-12b

12B text model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified
Recipe ready · unverifiedCommercial usetext

Xor 26B-A4B

26B text model.

Hardware estimate: NVIDIA RTX PRO 6000 Blackwell (96 GB) workstation · unverified

How it works

Inspect before the weights arrive

Installer plans use the pinned model revision and the constraints recorded for each local variant.

DetectRead CPU, GPU, memory, disk, and runtime capabilities.
RecommendPick a compatible precision and serving runtime.
DownloadFetch files from the official source and verify recorded checksums.
ServeStart a Jev-compatible API bound to loopback.
Self-testCheck the endpoint before sending real requests.

Local by default

Control the boundary.

Installer releases are signed, model downloads are checked against their recorded hashes, and the local API listens on loopback unless configured otherwise. The installer adds no usage telemetry.

cosign verify-blob --bundle SHA256SUMS.sigstore.json --certificate-identity-regexp '^https://github.com/fstandhartinger/decisionmodels-local/\.github/workflows/release\.yml@refs/tags/v[0-9][^/]*$' --certificate-oidc-issuer https://token.actions.githubusercontent.com SHA256SUMS

The bootstrap checks the signed manifest automatically. For a manual check, download SHA256SUMS and SHA256SUMS.sigstore.json from the GitHub release, then run this command.

Licence, separately

Model terms still apply.

Using the installer does not change a model's licence. Some weights limit commercial use or redistribution.

Questions

Before you install

Will every model run on a laptop?

No. Each page lists the recorded memory requirements and supported platforms. Use the plan command to check the machine you intend to use.

Where do model files come from?

The catalogue pins a model repository and revision. The installer uses that official source and checks recorded file hashes.

Does local installation include commercial rights?

No. Model licences are separate from the installer licence; review the model's own terms before commercial use.

Can I use my own GPU server?

Yes. The remote command connects to a machine you control over SSH. The remote host still needs a supported runtime and compatible hardware.

Run locally — Decision Models