Skip to content
Metatext Decide — typed decisions for your agents, with a probability on every answerGet early access
Metatext
Sign Up

local-jev-api

Docker recipes and ARM64 images for serving ten native decision checkpoints, such as Cloudflare Clef, behind /v1/systemone on DGX Spark.

Servicebeta
View on GitHub ↗

Overview

local-jev-api provides the recipe from the author's DGX Spark experiments for running native decision checkpoints behind a local /v1/systemone endpoint, plus prebuilt Linux ARM64 container images and a comparison of ten checkpoints. The default walkthrough uses Cloudflare Clef Flash, and a run script can launch any of the ten tested checkpoints by key. Each accepts a text state, questions and explicit answer choices; the README stresses this is a native decision-head API, not a vLLM chat server.

Two images are provided: a common runtime for nine checkpoints and a Clef Omni compatibility runtime. They target NVIDIA DGX Spark specifically, are around 20 GB unpacked, and exclude model weights, which download separately; Clef Flash weights take about 18 GiB and its peak host memory drop was about 36 GiB in the author's workload. At the time of writing, the container package was private, so anonymous pulls fail until it is made public. Image provenance and digests are recorded.

What the decision model decides

TypeQuestionAnswers
ChoiceWhich explicit answer choice fits the text state?—

How it works

  1. Clone the repository on a DGX Spark
  2. Authenticate to GHCR and pull the runtime image
  3. run.sh downloads the checkpoint and launches the server
  4. Send /v1/systemone requests
  5. Switch checkpoints with run.sh <model_key>

Strengths

  • One-command launch on DGX Spark
  • Ten checkpoints compared
  • Provenance recorded for images

Caveats

  • Images were private at the time of writing
  • DGX Spark / ARM64 only
  • Large memory and disk requirements

Getting started

git clone https://github.com/krisitown/local-jev-api.git && cd local-jev-api && ./run.sh

At a glance

Type
Service
Status
beta
Language
Python
Runtime
Docker on NVIDIA DGX Spark (ARM64)
Has code
yes

Use cases and integrations

How curated lists describe it

Frequently asked questions

Which hardware does it target?

NVIDIA DGX Spark on Linux ARM64.

Is it a chat server?

No, it serves native decision heads behind /v1/systemone.

Can I pull the images anonymously?

Not at the time of writing; the package was still private.

Related projects

  • cu-Jev

    CUDA-native engine that serves open Qwen3.5 checkpoints behind a Jev-compatible API, reading option logits without decoding.

    BenchmarkbetaC/CUDA
  • Laya-CoreML

    Core ML and Neural Engine port of Laya for Apple Silicon, about 5 ms per short decision on an M3 Max with energy benchmarks.

    BenchmarkbetaPython
  • systemone-lite

    Toy local Jev-shaped decision API on a 0.5B model that answers yes/no, choice and score questions by scoring option tokens.

    ServiceprototypePython
  • jiwo

    Small open decision models (0.8B and 4B, from Qwen3.5) with a server that speaks Jev's /v1/systemone format in one forward pass.

    ServicebetaPython
  • minicpm5-jev

    Serves MiniCPM5-2B-8bit on Apple Silicon as a Jev-compatible /v1/systemone service and reproduces its JevBench score.

    ModelprototypePython
  • reflex★ 166

    Small open decision model on a frozen Qwen3.5-4B serving /v1/systemone in about 200 ms, benchmarked against Jev on JevBench.

    ServiceresearchPython