Guides / Local installation

Run a local decision model in Python without a hosted API key

JevStyle runs downloaded weights on your machine to classify text, route requests and score fixed options. Install its Python package, choose a backend and make a request to the loopback server.

Reviewed 2 November 2026 · Python 3.10+ · Maintained by Chaoliang Yan

Install JevStyle and choose one local backend

Create a virtual environment first. On macOS or Linux:

python3 -m venv .venv
source .venv/bin/activate

On Windows, use py -m venv .venv, then activate it with .venv\Scripts\Activate.ps1 in PowerShell. Install one backend below. The first server start downloads the released weights; the commands run real model inference, not an agent simulation.

Apple silicon: small text-only model with MLX

python -m pip install -U "jev-style[mlx]"
jev-style serve --release 0.8b --backend mlx

CPU or CUDA: text-only baseline with PyTorch

python -m pip install -U "jev-style[torch]"
jev-style serve --release 0.8b --backend torch
# For a CPU-only run, add: --device cpu

Receipts or screenshots: 2B v4 with PyTorch

python -m pip install -U "jev-style[torch]"
jev-style serve --release 2b --backend torch

Choose PyTorch explicitly for images on a Mac. The MLX and GGUF builds, and all 0.8B builds, are text-only. The 2B v4 PyTorch weights total about 4.43 GB; download size is not total RAM. See the model and backend comparison.

Verify the local server and score a fixed choice

Leave the server running. Open the local Playground. In a second terminal, activate the same environment and run:

jev-style decide "I was charged twice for one order" \
  --url http://127.0.0.1:8765 \
  --choice "Which team?::billing,shipping,account,other"

The command prints the actual response from your local model. Your application validates the probabilities and keeps ambiguous or failed requests in a review path. It does not send inputs to the hosted JevStyle API. For GGUF, build the matching custom scorer; stock chat-generation commands do not expose JevStyle decision scores.

Local weights, hosted API and Jev skills are separate paths

Path	Where the decision runs	Credential	Source
Local JevStyle package	Downloaded JevStyle weights on your hardware	No hosted API key	Official source and CLI
JevStyle hosted API	Remote JevStyle service; text-only 2B v4	JevStyle account and key	Current service status and limits
Hugging Face demo	Remote demonstration service	No JevStyle hosted API key	2B text/image demo
TypeSafe/OpenRouter Jev access or agent simulation	That provider or the selected agent model	Provider-specific; not local JevStyle weights	Follow that service's own documentation

Training provenance and independence

Chaoliang Yan maintains JevStyle independently. 2B v4 was trained in part on soft labels distilled from TypeSafe Jev 1.14 through its public API. There is no TypeSafe affiliation or endorsement. The package does not bundle raw teacher-response datasets; that does not mean teacher outputs were absent from training. Consult the model card's training section for the release-specific sources.

Check fit before automating

Choose support-ticket classification, routing with a fallback, receipt and screenshot decisions, or fixed-choice scoring on Apple silicon. Evaluate on held-out examples. In the reproducible six-intent routing experiment, TF–IDF + logistic regression was more accurate and faster than JevStyle 0.8B; the development-selected JevStyle threshold missed its test-set error target. A successful install is not evidence of task accuracy.