Skip to content

Models

Small decision models you can run yourself

Each model scores a fixed set of options and returns a probability for each one. Weights are published under Apache-2.0.

Choose how to run Compare models See the evidence

Current models

Jev-Style 2B v4

2.2B · 2026-10-27

Decisions about text and images at 2B. Every option scored in one pass.

Transformers BF16 GGUF F16 / Q8_0 / Q4_K_M MLX BF16 / 8-bit

Jev-Style Cascade 9B

2B → 9B · 2026-10-03

Our 2B answers first; questions it is unsure of go to JevK5-9B by alibiserikbay.

Cascade config (no weights)

Jev-Style 0.8B v3

0.8B · 2026-09-24

Long-input, multilingual decisions in 0.8B.

Transformers BF16 GGUF F16 / Q8_0 / Q4_K_M MLX BF16 / 8-bit

MacJev 322M

322M · 2026-09-23

Decisions for local Mac agents.

PyTorch FP32 MLX FP32 GGUF F16

Earlier releases (3)

Jev-Style 2B v3

2B · 2026-09-27

Long-input decisions at 2B. Every option scored in one pass.

Transformers BF16 GGUF F16 / Q8_0 / Q4_K_M MLX BF16 / 8-bit

Decision 2B v2

2B · 2026-09-23

Calibrated English classification.

HF BF16 GGUF Q4_K_M / Q8_0 / BF16 MLX BF16

Decision 2B v1

2B · 2026-09-21

First release. Superseded by v3 (0.8B) and v4 (2B).

GGUF Q4_K_M / Q8_0 / BF16 MLX BF16

Three ways to run them

The hosted API and Hugging Face demo are separate services and receive what you send them. Local inference stays on your machine once the model is downloaded.

Hosted API

Runs 2B v4 for text only, with up to 8,000 input tokens. See the current shared-capacity limits.

Get API key API docs

Your own machine

Python 3.10+, inside a virtual environment. PyTorch reads text and images; MLX and GGUF read text.

python -m pip install "jev-style[torch]"
jev-style serve --release 2b --backend torch
# Open http://127.0.0.1:8765

In your browser

Try text and image examples on Hugging Face, with no installation or API key.