Jev alternatives

Von

The CPU answer, and the strongest open model on the independent board

Quick answer

Von is what to run when there is no GPU: a 395M encoder, ~1.5 GB on disk, under 15 ms per decision on a CPU. It is also the top open model on the independent 49-task benchmark at roughly 0.70–0.72 against Jev’s 0.966 — which is the honest headline of this entire category. Its recorded failure mode is specific: on unfamiliar domains it collapses to a single mode.

Von is a 395M ModernBERT-Large encoder with three decision heads — pick an option, answer yes or no, rate on a scale. It needs no graphics card at all, answers in under 15 ms on a CPU, and takes about 1.5 GB on disk. It also leads the open field on the one benchmark nobody involved in this ecosystem wrote: an independent 49-task classifier suite.

Train a decision model

A small non-autoregressive encoder with decision heads. No text generation at all — the classic classifier shape, rebuilt for natural-language options.

Search aliases

vonVonvon-sdkwfzyx/von

Key specs

Licence
Apache-2.0
Author
wfzyx
Backbone
ModernBERT-Large 395M + three decision heads
Size
~1.5 GB on disk
Latency
Under 15 ms per decision on a CPU — no GPU required
Wire format
von-sdk on npm for Node.js callers
Install
See repository

Von against Jev

The independent 49-task figure is the only number in this whole ecosystem produced by someone with no stake in any of the projects. Read it before any project’s self-reported score.

Von against JevVonJev
Independent 49-task benchmark~0.70–0.72 — the best open entrant0.966
Clean 20-bucket routing job~0.834 on a well-scoped taskNot published for this split
Messy 49-task suite~0.715 — the accuracy falls off a cliff outside its lane0.966
HardwareCPU, ~1.5 GB, no GPUHosted
LatencyUnder 15 ms locallyNetwork round trip
Recorded failure modeCollapses to a single mode on unfamiliar domainsSee published failure modes

When Von is the right choice

Choose Von when the deployment constraint is real: an edge box, a laptop, a CPU-only container, or an environment where sending state to a third party is not allowed. It is small enough to be boring, and it is the one open project that earns its place on a benchmark run by an outsider rather than by its own author. Point it at a narrow, well-defined decision — a routing bucket, a yes/no gate — and it is fast and adequate.

When to walk away

Do not use Von for open-ended or unfamiliar domains. Its independently recorded failure mode is collapsing to a single answer when the domain is not the kind it was pointed at, and its own numbers fall from roughly 0.834 on a clean routing job to about 0.715 on a messy 49-task suite. If your inputs are unpredictable, that gap is the whole story, and a cascade up to a stronger model is cheaper than a wrong decision at scale.

Adopting Von in three steps

01Run it on the machine you already have — CPU, 1.5 GB, no accelerator needed. There is a Node.js SDK if your caller is JavaScript.
02Scope the decision tightly. Von is strongest on a well-bounded routing or yes/no task and weakest when the input distribution is wide.
03Put a confidence gate in front of any consequential action, and measure your own held-out set before trusting the model card.
bash / model + sdk
# Von needs no GPU: 395M params, ~1.5 GB on disk.
# Each decision returns in under 15 ms on a CPU.

git clone https://github.com/wfzyx/von.git && cd von
# install per the README, then serve or call it in-process

# A Node.js caller can use the published SDK:
npm install von-sdk
#   import { decide } from 'von-sdk'
#   const r = await decide({
#     state: 'We were billed twice for March.',
#     question: 'Which team should handle this?',
#     options: ['billing', 'technical', 'sales'],
#   })
#   if (r.confidence < 0.85) escalateToHuman(r)
Under 15 ms per decision changes the architecture: you can gate every single action rather than sampling, and the whole model fits in memory next to your application.

Von questions people actually ask

Can Von really replace Jev?

For a narrow, well-defined decision on hardware without a GPU, it can be good enough. As a general replacement, no: the independent 49-task benchmark puts the best open entrant around 0.70–0.72 against Jev’s 0.966. Nobody in this ecosystem is claiming otherwise with clean evidence.

Why is Von rated above Laya when it has far fewer stars?

Because stars measure attention, not accuracy. On the independent 49-task benchmark Von leads the open field while Laya sits at 0.583. Laya has upwards of nineteen thousand stars; Von had a few hundred. That mismatch is the single most useful thing to understand about this category.

What does Von fail at?

Unfamiliar domains. The benchmark records its failure mode as collapsing to a single mode when the input does not resemble what it was pointed at, and the project’s own numbers drop from about 0.834 on a clean 20-bucket routing job to roughly 0.715 on a messy 49-task suite. Always measure your own distribution.

Do I need any special hardware?

No. That is the point. Von is a 395M encoder, about 1.5 GB on disk, and it answers in under 15 ms on a CPU, so it runs on a laptop, an edge device or a CPU-only container — the cheapest deployment story in this comparison.

Sources

Every comparison number on this page is a third-party published figure or a read of a public repository — not a benchmark we ran. We have not tested TypeSafe Jev itself, and its customer agreement forbids using its outputs to build similar products.