Dev-0.4B
399M ModernBERT-large decision model with one universal choice head answering Choice, Noul and Score in a single forward pass.
Overview
Dev (dev-0.4b) is an open 399M-parameter bidirectional decision model for developer tools and coding agents, following Jev and Jared Palmer's Kev. It uses a ModernBERT-large encoder with a single universal choice head that computes decision probabilities from pooled candidate spans in one forward pass, instead of constrained decoding on large generative models. The README reports about 28 ms on Apple Silicon and about 10 ms on CUDA.
The README compares it with kev-0.5b on held-out public benchmarks at a 0.5 threshold, reporting 91.33% versus 86.0% on Banking77, 85.20% versus 75.3% on BoolQ and 62.67% versus 55.3% on Yelp 5-star sentiment, and a modest code-retrieval reranking gain over BM25 on CodeSearchNet. According to a list summary, every figure reconciles to an evaluation JSON shipped in the repository. Weights are on Hugging Face.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which candidate option fits (e.g. 77-way intent routing)? | — |
| Yes / no | Is this statement true given the passage? | yes · no |
| Score | Which star rating fits this review? | 1 · 2 · 3 · 4 · 5 |
How it works
- Load dev-0.4b from Hugging Face
- Encode state and candidate spans together
- Choice head scores each candidate
- Apply a threshold or argmax
Strengths
- Small and fast (tens of milliseconds)
- Single head for all three primitives
- Evaluation JSON shipped with the repo
Caveats
- Comparisons are against a similarly small model
- Encoder context limits apply
Getting started
Download the weights from huggingface.co/mpnikhil/dev-0.4b and follow the repository's inference example.At a glance
- Type
- Model
- Status
- research
- Language
- Python
- Runtime
- PyTorch (MPS / CUDA)
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
A 399M bidirectional encoder with one universal choice head, answering Choice, Noul and Score in a single forward pass. Every README figure reconciles to an evaluation JSON shipped in the repository.
MrJev/awesome-jev
Frequently asked questions
What architecture is Dev?
A ModernBERT-large encoder with a single universal choice head.
How fast is it?
About 28 ms on Apple Silicon MPS and about 10 ms on CUDA, per the README.
How does it compare with Kev?
The README reports higher scores than kev-0.5b on Banking77, BoolQ and Yelp.
Related projects
Mangaba AI's local, calibrated typed-decision model for Portuguese text, trained to optimize the Brier score; proprietary since v5.
ProjectproductionPython- blink★ 22
Open one-pass typed-decision model with a dependency-free C99 runtime that also builds to a 66 KB WebAssembly module.
ServicebetaC - Foq★ 11
Local open alternative to Jev using a 1.58-bit ternary 8B model to answer booleans, choices, scores and Pydantic extraction in ~25 ms.
ProjectbetaPython Runs the open Laya decision model entirely in a browser tab with WebGPU and ONNX Runtime Web: no server, no API key.
ServicebetaTypeScriptllama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++Three open 4B decision models for English, Turkish and German that serve the Jev /v1/systemone API on a CPU in about 3 GB of RAM.
ServicebetaPython