alternatives / choosing a replacement

Jev alternatives, ranked by something better than stars

Logit readout, dedicated encoders, adapted LLMs and diffusion — four routes that replace different parts of Jev. Every number carries a source, every project carries the failure mode its own authors admitted, and the board-version problem nobody else mentions is on this page too.

Quick answer: which one should you use?

Shortest migration with no code changes: Kev — same /v1/systemone contract. No GPU at all: Von — 395M, under 15 ms on a CPU. Reuse an open model you already host and train nothing: SemIf or AnyJev. Best result after fine-tuning on your own labels: Laya, and prefer its successor cbjev (mind the GPL-3.0). Tight control loops: NanoJev. Image state or diffusion research: djev and OpenJev.

Star counts and latency figures are a snapshot from 2026-09-23/25 and go stale within days. · Projects covered: 30

Four factions that replace different things

Within a week of Jev shipping, the community had already sorted the open reproductions into four approaches (BlockBeats). Keeping them apart is the only way to see why these "alternatives" cannot be compared to each other directly.

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.

  • Laya
  • Von
  • NanoJev
  • cbjev
  • jevlike
  • JevK5

Read the logits

No new weights. Intercept the model right before it answers and read the probability of each option instead of letting it write.

  • SemIf
  • AnyJev
  • Verdict
  • litjev
  • simple-jev
  • fastjev

Adapt a large model

Keep the LLM as the backbone and bolt a decision structure onto it — LoRA adapters or a pointer head — so one prefill answers every question.

  • Kev
  • Nimble
  • decider (Mapika)
  • Decitron (决策机)
  • Winnow-12B
  • Open-Jev (Zefan Cai)

Fill answers with diffusion

Leave the answer slots blank and fill them all at once, the way a diffusion model completes an image, instead of decoding token by token.

  • djev
  • OpenJev (razorback16)
  • LocalJev
  • openjev-sglang
  • DiffusionGemma
  • Lichen

The full comparison table

Licence, backbone, latency, stars and the case each one suits. Projects with a page here link internally; the rest link straight to the source repository.

ProjectLicenceBackbone & sizeLatencyStarsBest for
LayaConvai InnovationsApache-2.0ModernBERT-large 421M · mmBERT-base 322M (multilingual)32.8–39.5 ms/question on a T4 · 7.2 ms for ten batched · other measurements put single-question CPU inference at 193–464 ms19,301A fast multilingual base you fine-tune on your own labels — not a zero-shot replacement.
KevJared PalmerApache-2.0Qwen3.5 0.8B / 4B / 9B + rank-16 LoRA + pointer head47 ms on MLX (repeated state), 77 ms fresh · 12–26 ms on an H1005,385Replacing a hosted Jev call without touching your code — it serves the same /v1/systemone contract.
SemIfTheo LeeMITFrozen Qwen3.5-4B — no fine-tune at all1.023 s for 21 binary criteria · 5.21× faster than generating the same answers as JSON4,023Reusing an open-weight model you already host, with zero training and zero output tokens.
VonwfzyxApache-2.0ModernBERT-Large 395M encoder + three decision headsUnder 15 ms per decision on a CPU — no GPU at all546Hardware-free deployment: a laptop, an old server, an edge box. Also the top open model on the independent 49-task benchmark.
NanoJevTianyuCodingsNot stated in the repository; third parties list MITQwen3-0.6B + parallel decision headsCUDA only — the inference script has no Apple Silicon path2,086Tight control loops — the only open project that actually beats Jev at a task (ViZDoom Basic 128/128 vs 56/128).
cbjevtomek7667GPL-3.0-or-laterModernBERT / mmBERT, fine-tuned from Laya3.0 ms for one question · 11.4 ms for ten questions over a 500-token document—A faster, better-calibrated Laya successor that keeps the same wire format. Note the GPL-3.0 licence.
AnyJevJiamu (Morris) Zhang · Nokia Applied ResearchApache-2.0Any open LLM — a training-free debiasing layerOne prefill per call; no generation—Killing option-order bias and label bias without a single label, then calibrating with 100–500 labels.
djevMaisaApache-2.0 per the repositoryDiffusionGemma 26B-A4B read as a canvas; no weights shippedJevBench 74.3 → 73.0 → 52.23 across the three board versions, with no new measurements · 76.87 ms p50 on the older quantized config—The diffusion route, and one of two entries cited for image state. Evidence is thin — read the repositories and check which board a score came from.
jevlikevinnylarougeMITByte encoder, or any frozen Hugging Face encoderCPU / MPS / CUDA — four commands to train1,255A trainer, not a model. Bring your own option list and labels, train on a laptop.
VerdictManavarya09 (one of three unrelated projects using the name)Apache-2.0multilingual-e5-small 118M (PyPI verdictml)0.5 ms/example batched on an M5 CPU · under 2 ms single with ONNX int8—Honest probabilities: temperature scaling plus conformal abstention with a coverage guarantee, and option order that cannot change the answer. Weigh its self-reported 0.77 against Jev against a JevBench v1.4.2 row of 5.69.
OpenJev (razorback16)razorback16 — note: "OpenJev" names at least seven unrelated projectsApache-2.0DiffusionGemma 26B-A4B with a one-step structured read~94 ms median on an RTX PRO 600050Cited as the only open implementation that answers questions about images, and the reference the other bridges copy. Always name the author: bare "OpenJev" is unusable.
LocalJevGitHub NextSee repositoryDiffusionGemma 26B-A4B through an OpenAI-compatible endpointDepends on oMLX; a TypeScript/Bun bridge—A portable bridge when your runner lacks the diffusion primitives. Its README is explicit: wire-compatible, not mathematically equivalent.
openjev-sglangekzhangSee repositoryQwen3.6-35B-A3B served on SGLangB200-class hardware—Serving a large open decision model at high throughput when you have datacenter GPUs.
litjevzhengxuyuApache-2.0Open weights (Qwen and others), option logitsSee repository13A small, readable /v1/systemone server to study or fork if you want to self-host the whole contract. It publishes no Jev-vs numbers of its own.
simple-jevfeatherless-aiMITAny OpenAI-compatible endpointSee repository168The fastest zero-training experiment: point it at a vLLM or Hugging Face endpoint and read options off the logits.
fastjevchengyongruSee repositoryA SemIf forkSee repository—A leaner SemIf if you want self-hosted semantic decisions with less scaffolding.
JevK5allebeeApache-2.0Open weights, one forward passSee repository—Apache-2.0 open weights when you want to inspect or fully own the decision model.
poorjevrupeshpoojary9See repositoryRuns locally with no API keyECE 0.170 → 0.071 after calibration—Offline, no-waitlist experiments where calibration matters more than raw accuracy.
NimbleBespokeSee repositoryQwen3.5-9B with contrastive post-trainingPicked the reference answer 90.1% of the time on 324 held-out samples (Jev: 93.2%)—Teams that care about the training recipe: data built so that flipping one fact flips the correct answer. The repo ships no LICENSE file — the Apache-2.0 claim rests on the model card alone.
decider (Mapika)Mark MarosiApache-2.0Qwen3.5 2B / 4B / 35B-A3B-Base + one-pass letter-slot readout32.3 ms median on a B300 · 3.2 ms with CUDA graphs—Top of the current JevBench board (v1.4.2, 64.13). Note the widely repeated "33 ms" figure is Laya on a T4, not decider.
GLiNER2.5-DecideFastinoSee source340M per the vendor (some cards say 205M — the conflict is unresolved)0.698 macro accuracy on the independent 49-task benchmark—Routing and extraction-shaped decisions where an existing GLiNER pipeline is already in place.
sokudan (即断)GeneLabApache-2.0 code · MIT backbonesbintuitions/modernbert-ja-310m with cross-attentionDay-1 results published without releasing weights (bool still below the majority baseline)—Japanese-native decisions. Its published day-1 run already beats every baseline on the score (ordinal) task.
DiffusionGemmaGoogleGemma terms26B-A4B diffusion language model27.4 decisions/s in batch routing—An engine rather than an alternative — the model several of the projects above are built on.
Decitron (决策机)中科闻歌 / Zhongke WengeProprietaryA general-purpose decision LLMNot comparable—NOT an open Jev alternative. The axis is "instant judgement vs future deduction" — a different category, often confused with a clone.
JevForgezwliJayMITQwen3.5 0.8B / 0.6B with an end-to-end synthesis-train-calibrate-serve toolkitTest noul 0.826 / 0.776 vs Jev 0.910 · out-of-distribution noul 0.860 / 0.769 vs Jev 0.825—The only entrant we found that beats Jev on any axis out of distribution — while losing in distribution. Worth independent verification before repeating.
Winnow-12B—Apache-2.012B, 16 GB GPU198/231 on one JevBench subset — level with Jev’s 85.71% there—A mid-size option that reaches parity with Jev on a public subset. Check the sealed-accuracy column before treating that as a real win.
Open-Jev (Zefan Cai)Zefan CaiMIT codeLoRA + scalar head + calibration temperature on pinned Qwen3.5 2B / 9B / 27B85 ms against Jev’s 295 ms—Reports 197/231 against Jev’s 200/231 on a public JevBench run — the closest public-scoring open result, on a narrow slice.
LichenMushroom-SystemsMITDocker / llama.cpp over GGUF (gemma-4-26B-A4B)93 ms · 207/231 against Jev’s 200/231, p = 0.09 (not significant)—A containerised way to stand up a decision endpoint. It even reports that its own lead over Jev is not statistically significant.
Cygnetblockbrain (Nood Co)MIT shim / Apache-2.0 weightsFrozen gemma-4-12B-it plus an MIT shim, temperature 3.4 — no fine-tune50–72 ms · JevBench v1.4.2 #4 at 61.76—Scores above decider-4b v2 on the intelligence sub-score while training nothing. Its sealed accuracy is 33.8% against 34.7% for decider.
HopperHopitAIApache-2.0 code, but "do not use this adapter commercially"LoRA r16 on Qwen3.5-4B, ~9 GB, 16 GB GPU minimum44–57 ms · JevBench v1.4.2 #5 at 59.43—Calibration sub-score of 79.06, above Jev’s 76.34. The licence text forbids commercial use, so read it before benchmarking it internally.

Rank by the independent benchmark, not by stars

This is the only benchmark in the ecosystem run by someone with no stake in any of the projects. It also records what each system gets wrong.

Systemmacro accuracyRecorded failure mode
TypeSafe Jev (hosted)0.966
Von (395M, open)0.704Collapses to a single mode on unfamiliar domains
GLiNER2 (~300M, open)0.698Over-triggers on keywords
Laya (421M, open)0.583Compresses rating scales

49 tasks, 869 cases across compliance, triage, legal, DevOps, linguistics and safety. v2 macro accuracy. jabr/classifier-benchmark

One project, three scores: the JevBench board problem

The same systems were re-scored twice with no new measurements. Whenever a post says something "beats Jev", ask which board it is quoting.

BoardLeaderJevdjevNote
v1.2.8-eraJev 75.375.3 (#1)74.3 (#3)The board most blog posts still quote.
v1.3.0Jev 74.474.4 (#1)73.0 (#3)Re-scored with chance-corrected intelligence and a near-chance penalty — no new measurements.
v1.4.2decider-4b v2 64.1363.29 (#2)52.23 (#8)308 sealed decisions and a harmonic mean with a generalisation gate. Jev is no longer first.

Name collisions

In this ecosystem a bare project name is unusable — the same name can mean entirely different things.

  • OpenJev × 7

    At least seven unrelated repositories: SemIf’s former name, Zefan Cai’s Open-Jev, S1LV3RJ1NX, razorback16, SiliconLabAI (a Next.js playground, not a model), and others. Always name the author.

  • Verdict × 3

    Manavarya09/verdict (118M, PyPI verdictml), openJev-verdict-2.0 (149.6M ModernBERT + GLiClass) and the JevBench row "verdict-small" are three different projects.

Jev alternatives: frequently asked

Is there a real open source replacement for Jev?

At the interface level, yes. At the accuracy level, not yet. Kev, cbjev and others implement the same /v1/systemone contract, so one base-URL change swaps the call. But on the independent 49-task benchmark the strongest open model scores roughly 0.70–0.72 against Jev’s 0.966. Open projects already win on cost, latency and data residency; they lose on zero-shot, out-of-domain quality.

Why do different posts report wildly different scores for the same project?

Because JevBench has three board versions. The same systems were scored in the v1.2.8 era, then on v1.3.0, then on v1.4.2, with no new measurements in between: djev moved from 74.3 to 52.23, one project from 67.0 to 40.6, another from 62.4 to 27.4. Always ask which board a number came from.

Which project is "OpenJev"?

At least seven unrelated repositories use that name: SemIf’s former name, Zefan Cai’s Open-Jev, S1LV3RJ1NX, razorback16, SiliconLabAI (actually a Next.js playground, not a model) and more. Verdict collides three ways. That is why this site always writes author/repository rather than a bare project name.

Can I use Laya zero-shot?

No. Laya’s own README reports base checkpoints at 0.362 and 0.342 on typed decisions against a 0.318 random baseline and a 0.461 majority baseline; the widely quoted 0.766 comes from a checkpoint fine-tuned on that benchmark’s training split. A 300-item Japanese test also found severe option-position bias in its ordinal output — the first option was chosen 0 or 1 times out of 300. Use it as a fine-tuning base.

What is the real production risk with any of these?

Not accuracy — looking adequate. Position bias, confident errors, and benchmarks that overlap their own training data all produce models that demo well and fail quietly. Before adopting any alternative, label a few hundred examples from your own workload and measure three things: whether the confidence is trustworthy (ECE), whether reordering options changes the answer, and whether low-confidence cases really are harder.

How current is this data?

Star counts and latency figures are a snapshot from 2026-09-23/25; benchmark scores are dated by their own sources. This ecosystem changes weekly — the first open reproductions appeared within 24 hours of Jev shipping, six clones were counted in two days, and more than a dozen within a week. Re-verify any specific number before you rely on it.

We have not benchmarked TypeSafe Jev ourselves. TypeSafe’s Master Customer Agreement 2.3(b) forbids using the service or its outputs to build similar products, which is why clean head-to-head runs barely exist in this ecosystem. Every number on this page is a third-party published figure or a read of a public repository, and should be treated as a hypothesis rather than a result.