Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › JEV family

JEV-27B

Hugging Face ↗ Follow changes

autotrust Apache 2.0 out 30 days ago

26.9B parameters, text in, labels out, 256K tokens of context.

Verified 2026-10-25 Rank #0 in Decision models · score 0 method v1 Sources 3 Downloads 30d 384.5k · #186 of 619 tracked

Can I run it?16GB at INT4Estimatefits a 24 GB cardmethod
Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?384.5kdownloads, 30 days#186 of 619 we track · top 30%−5 likes, downloads +90% this week85 likesHugging Face2026-10-24

Worth reading

Picked by rule from 15 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 1,989 HN pts

Introducing System One Models and Jev

TypeSafe AI releases Jev, a System One Model designed for fast, structured automation decisions.

  • Jev outputs all probabilities in parallel instead of autoregressively generating by token.
  • The model never makes type errors because possible outputs and structure are defined in advance.
typesafe.ai 2026-09-15 1,989 HN pts · 520 comments 2.2k words about the JEV family family
Latest · 26 days ago

Jeeves. Reasoning improves Jev-like decision models

Jeeves is a 9B Jev-like model based on Qwen3.5-9B that reasons only when unsure, trained on public data to improve decision accuracy.

  • Jeeves v2 was trained on chains of at most 1,024 tokens, so lower caps cut most of its chains.
  • The model uses rare tokens like "<state>" and "<decide>" in the prompt, as using plain text like "State" worsened performance.
github.com 2026-09-29 242 HN pts · 95 comments 2.4k words about the JEV family family

15 best of 15 tracked

The 2 picks above, plus these 13. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

  1. CodeJeff – Jev-compatible 0.8B decision models, trained at home, ~30 msgithub.com · 2026-09-28 · 575 HN pts
  2. CodeKev: Tiny Jev-like family of decision models built on top of Qwen3.5github.com · 2026-09-21 · 463 HN pts
  3. ArticleOllaya – Ollama for open-source, Jev-style decision modelsollaya.dev · 2026-09-25 · 618 HN pts
  4. ArticleLanguage models for text classification: From bag-of-words to Jevmagazine.sebastianraschka.com · 2026-09-29 · 213 HN pts
  5. CodeI turned Jev into a (lousy) chatbotgithub.com · 2026-09-20 · 177 HN pts
  6. ArticleDecision models like Jev don't beat LLM-as-a-judge or traditional classifiersdevelopers.redhat.com · 2026-10-02 · 165 HN pts
  7. CodeReverse-engineered Jev-like modelgithub.com · 2026-09-16 · 169 HN pts
  8. ArticleTurning GLM-5.3-Flash into a Jev-like decision modelprivatemode.ai · 2026-09-26 · 138 HN pts
  9. ArticleA single function Jev-like wrapper for LLMs, including vision modelsallanrbo.blogspot.com · 2026-09-26
  10. ArticleShow HN: JevBench, a reproducible benchmark for typed decision modelsbenchmarkheaven.com · 2026-09-22
  11. ArticleShow HN: Jevstiller – Distill Jev into a local model, with a disagreement boundjevstiller.pages.dev · 2026-09-29
  12. DiscussionOpen-sourced jev architecture last year with model,paper and datasetnews.ycombinator.com · 2026-09-17
  13. ArticleWhat Is RLCD? The Secret Behind Jevdi-zhang-llm.github.io · 2026-09-24

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 2 quantized builds and 10 Spaces.

  • As a service0no API provider listed
  • On own hardware2quantized builds
  • Built on00 fine-tunes · 0 adapters · 0 merges
  • In apps10Hugging Face Spaces

Top Spaces running it: multimodalart/jev-decision-index ♥ 802 · autotrust/jev-9b-decision-demo ♥ 62 · autotrust/JEV-27B-Demo ♥ 30 · hugging-apps/jev-27b-vl ♥ 3

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 64 GB Mac (unified memory) at INT4

About 16 GB at INT4, 32 GB at FP8, 65 GB at BF16: weights only, so leave room for the KV cache.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 16 GB of weights (estimate; KV cache and activations extra)INT416 GBFP8: about 32 GB of weights (estimate; KV cache and activations extra)FP832 GBBF16: about 65 GB of weights (estimate; KV cache and activations extra)BF1665 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 16 GB of weights (estimate; KV cache and activations extra)INT416 GBFP8: about 32 GB of weights (estimate; KV cache and activations extra)FP832 GBBF16: about 65 GB of weights (estimate; KV cache and activations extra)BF1665 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

Coverage peaked at launch and is still arriving

Teal is the weights (2026-09-25), amber triangles are 10 dated articles (each links to it), grey ticks are repo commits.

Oct2026-10-11: 6 repo commits2026-10-10: 4 repo commits2026-10-02: 2 repo commits2026-10-01: 2 repo commits2026-09-30: 1 repo commit2026-09-26: 1 repo commitJeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms · 2026-09-28Introducing System One Models and Jev · 2026-09-15Kev: Tiny Jev-like family of decision models built on top of Qwen3.5 · 2026-09-21Jeeves. Reasoning improves Jev-like decision models · 2026-09-29Ollaya – Ollama for open-source, Jev-style decision models · 2026-09-25Language models for text classification: From bag-of-words to Jev · 2026-09-29I turned Jev into a (lousy) chatbot · 2026-09-20Decision models like Jev don't beat LLM-as-a-judge or traditional classifiers · 2026-10-02Reverse-engineered Jev-like model · 2026-09-16Turning GLM-5.3-Flash into a Jev-like decision model · 2026-09-26Weights 2026-09-25 · 2026-09-25Weights 2026-09-25now
Oct2026-10-11: 6 repo commits2026-10-10: 4 repo commits2026-10-02: 2 repo commits2026-10-01: 2 repo commits2026-09-30: 1 repo commit2026-09-26: 1 repo commitJeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms · 2026-09-28Introducing System One Models and Jev · 2026-09-15Kev: Tiny Jev-like family of decision models built on top of Qwen3.5 · 2026-09-21Jeeves. Reasoning improves Jev-like decision models · 2026-09-29Ollaya – Ollama for open-source, Jev-style decision models · 2026-09-25Language models for text classification: From bag-of-words to Jev · 2026-09-29I turned Jev into a (lousy) chatbot · 2026-09-20Decision models like Jev don't beat LLM-as-a-judge or traditional classifiers · 2026-10-02Reverse-engineered Jev-like model · 2026-09-16Turning GLM-5.3-Flash into a Jev-like decision model · 2026-09-26Weights 2026-09-25 · 2026-09-25Weights 2026-09-25now

What people built on it

2 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

2 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

2 quantized (Hugging Face model tree)2 quantized
2 quantized (Hugging Face model tree)2 quantized
GGUF prithivMLmods/JEV-27B-GGUF for llama.cpp, Ollama, LM Studio 2.1k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

Benchmarks

15 rows

Benchmarks 15 rows

measured third-party leaderboard rowauthor-reported number from the model card or eval files (15)No confidence intervals unless the source gave them; settings differ between rows, so compare within a benchmark, not across. Date = the benchmark's own date when it gives one, else when we read it.
BenchmarkValueSetting / notesReported byDate
author-reported Choice top-1 agreement with Jev, Jev-labelled rows90.5 % percentSystem 1model author2026-10-14
author-reported Expected calibration error0 otherSystem 1model author2026-10-14
author-reported JevBench88.7 percent—model author2026-10-14
author-reported KL to the programmatic labels of task families never seen in training (Open-Jev OOD split)0.1 otherSystem 1model author2026-10-14
author-reported Kev83.75 percent—model author2026-10-14
author-reported MASSIVE-en87.71 percent—model author2026-10-14
author-reported Mean KL divergence from TypeSafe Jev 1.13's distributions, Jev-labelled rows, 0 = identical≈ 0.017 otherSystem 1model author2026-10-14
author-reported Mean KL to all test targets (Jev, programmatic and placeholder labels)0.02 otherSystem 1model author2026-10-14
author-reported Nimble92.91 percent—model author2026-10-14
author-reported OpenJev text73.89 percent—model author2026-10-14
author-reported Rating error (`score`, 0–5 scale), mean absolute error of the expected rating0.1 otherSystem 1model author2026-10-14
author-reported Six-group mean84.07 percent—model author2026-10-14
author-reported Terminal-Bench · Write code with agents96 % of the teacher percentSystem 1model author2026-10-14
author-reported VitaminC77.46 percent—model author2026-10-14
author-reported Yes/no AUROC (`noul`), Jev-labelled rows0.99 otherSystem 1model author2026-10-14

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

21 fields

Listed in Decision models

Facts

Pipeline tagtext-classification
Librarytransformers
ArchitecturesQwen3_5ForCausalLM
Model typeqwen3_5_text
Dtypebfloat16
Params by dtypeBF16: 26,895,998,464
Size bandL
MoEno
Context length262,144 (config.json max_position_embeddings)
Languagesen
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen3_5ForCausalLM · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3_5ForCausalLM · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for JEV-27B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)65 GB · Single high-end GPU (<80 GB)
FP8 (1 byte/param)32 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)16 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra.

Decision facts

catalog

Decision-model catalog curated

How it was madeadapters (108.9M trained params) distilled from Jev 1.13 outputs; ~9.2 B200-hours
Lineage bucketJ - distilled from Jev outputs
Lineage evidencemodel card: "student of TypeSafe Jev 1.13"; dataset SargeDev/jev-distill-corpus-v3
Base modelQwen/Qwen3.8-27B
Hosted price (input)self-hosted/M
Notes25,376 of 29,955 test targets are Jev distributions
Sourcehttps://huggingface.co/autotrust/JEV-27B

All variants

2 variants

Lineage

Adapter Family: autotrust/JEV 2 derivatives on the Hub

Base model: Qwen/Qwen3.8-27B

quantized: 2

Top variants (2 shown, ranked by downloads)

2 rows
Variant Relation Quant Params Band License Author Likes Downloads
prithivMLmods/JEV-27B-GGUF quantized GGUF — — apache-2.0 prithivMLmods 3 2.1k
webmp3/Sakura-Humanlike-Chat-2.0-JEV-AutoRound-GGUF quantized GGUF — — apache-2.0 webmp3 0 1.9k

Repository changelog

16 commits

Repository changelog 16 commits

DateCommitSHA
2026-10-11 Model card: AutoTrust/GuruSearch recommendation & search live demo (news.guru.so) 14b4252
2026-10-11 Banner: AutoTrust/GuruSearch recommendation & search live demo 2dcbb0d
2026-10-11 Model card: autotrust/JEV recommendation & search live demo (news.guru.so) 8e3842f
2026-10-11 Banner: autotrust/JEV recommendation & search live demo 8f2286d
2026-10-11 Model card: news search powered by autotrust/JEV (news.guru.so) 03fc6dc
2026-10-11 Model card: JEV-powered news search is live (news.guru.so) 4469ae4
2026-10-10 Model card: fix the ModelScope homepage sentence (real-time sync) ff9b67b
2026-10-10 Model card: ModelScope section — links only (no recommendation or install lines) 975a3e3
2026-10-10 Model card: ModelScope section heading without '(China)' 10eecb4
2026-10-10 Model card: ModelScope (China) download links — official AutoTrust-Labs homepage, real-time sync with this repository 0f7a4d3
2026-10-02 serve_decide.py: System 1 only by default for every model (thinking is opt-in) 51740a8
2026-10-02 serve_decide.py: one file for JEV-27B and GEV; optional adaptive thinking (thinking=auto|on, reasoning_effort, think_budget; off by default for JEV-27B), answer read-out that also works with speculative decoding e6b6a72
2026-10-01 Model card: "New (1 October 2026)" summary at the top 962701f
2026-10-01 System 1 over HTTP (POST /v1/decide, serve_decide.py), choice up to 256 options, prompt guide, 256K context test c917c6d
2026-09-30 Update README.md 924a8c7
2026-09-26 Publish JEV-27B model and evaluation results 9b1f6fd

Articles and coverage

15 pieces

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Analysis10

Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 ms
github.com2026-09-28authorityHN front pageabout the JEV familyHN 575 pts · 232 comments3,771 words
Introducing System One Models and Jev
typesafe.ai2026-09-15HN front pageabout the JEV familyHN 1,989 pts · 520 comments2,189 words
TypeSafe AI releases Jev, a System One Model designed for fast, structured automation decisions. Jev achieves similar intelligence to existing LLMs while being two orders of magnitude faster and more efficient, with end-to-end response times of 70ms-500ms compared to 3 to 329 seconds for frontier models. The model costs $0.042 per million input tokens and provides 40x-200x speedups for System One shaped queries by outputting calibrated probabilities in parallel rather than generating strings autoregressively.written, reviewed, cited
  • Jev outputs all probabilities in parallel instead of autoregressively generating by token.
  • The model never makes type errors because possible outputs and structure are defined in advance.
  • Jev supports a cardinality up to 255, requiring a two-stage system for higher choices.
Kev: Tiny Jev-like family of decision models built on top of Qwen3.5
github.com2026-09-21authorityHN front pageabout the JEV familyHN 463 pts · 210 comments6,746 words
Kev is a family of small decision models built on Qwen3.5 and Qwen3.8 that you can train and run locally. The API matches TypeSafe's System One, so you can point their Python SDK at your local server. Models use a small adapter on a frozen base, except Kev-27B which ships as 51 GB of full weights. The 27B model achieves 0.851 accuracy on new sources.written, reviewed, cited
  • The Kev-27B model requires 96–128 GB of RAM on Macs, a requirement marked as expected but not measured.
  • An explicitly requested accelerator fails instead of falling back if the probe times out or crashes.
Jeeves. Reasoning improves Jev-like decision models
github.com2026-09-29authorityHN front pageabout the JEV familyHN 242 pts · 95 comments2,362 words
Jeeves is a 9B Jev-like model based on Qwen3.5-9B that reasons only when unsure, trained on public data to improve decision accuracy. It beats Kev-9B on test data it was never trained on with 0.905 vs 0.822 accuracy. Inference runs on one H100 with --precision fp8 at about 20 ms per request without thinking and a 1.2 s median with reasoning.written, reviewed, cited
  • Jeeves v2 was trained on chains of at most 1,024 tokens, so lower caps cut most of its chains.
  • The model uses rare tokens like "<state>" and "<decide>" in the prompt, as using plain text like "State" worsened performance.
  • Every training chain reaches the right answer, so answers after reasoning come out overconfident and need the larger temperature.
Ollaya – Ollama for open-source, Jev-style decision models
ollaya.dev2026-09-25HN front pageabout the JEV familyHN 618 pts · 145 comments1,022 words
Ollaya runs open decision models locally with Apache-2.0 licensing, serving as a drop-in, Jev-compatible replacement for TypeSafe's hosted decision models. It achieves 0.824 accuracy on 3,880 questions using BF16 weights, while Ollama 0.40.6 runs 5 to 8 times faster via Q8_0 GGUFs. The platform supports 25 model families across macOS, Windows, Linux, and Docker, with NVIDIA GPUs requiring driver R525 or newer.written, reviewed, cited
  • The decider model by Mapika reads answers from option-letter logits in one forward pass.
  • The gliclass model scores all options in one pass, causing CPU cost to grow with option count.
  • The kev model uses a LoRA on a Qwen3.5 base plus a pointer head that scores every option at its own span.
Language models for text classification: From bag-of-words to Jev
magazine.sebastianraschka.com2026-09-29authorityHN front pageabout the JEV familyHN 213 pts · 10 comments7,738 words
I turned Jev into a (lousy) chatbot
github.com2026-09-20authorityHN front pageabout the JEV familyHN 177 pts · 50 comments1,006 words
Decision models like Jev don't beat LLM-as-a-judge or traditional classifiers
developers.redhat.com2026-10-02HN front pageabout the JEV familyHN 165 pts · 75 comments4,738 words
Reverse-engineered Jev-like model
github.com2026-09-16authorityHN front pageabout the JEV familyHN 169 pts · 24 comments1,019 words
Turning GLM-5.3-Flash into a Jev-like decision model
privatemode.ai2026-09-26HN front pageabout the JEV familyHN 138 pts · 59 comments2,447 words
All coverage tracked for this model (15) — append-only, nothing is ever removed
TitleSourcePublishedFirst seenSignals
Decision models like Jev don't beat LLM-as-a-judge or traditional classifiersdevelopers.redhat.com2026-10-022026-10-07hn
Show HN: Jevstiller – Distill Jev into a local model, with a disagreement boundjevstiller.pages.dev2026-09-292026-10-07hn
Jeeves. Reasoning improves Jev-like decision modelsgithub.com2026-09-292026-10-07hn
Language models for text classification: From bag-of-words to Jevmagazine.sebastianraschka.com2026-09-292026-10-07hn
Jeff – Jev-compatible 0.8B decision models, trained at home, ~30 msgithub.com2026-09-282026-10-07hn
Turning GLM-5.3-Flash into a Jev-like decision modelprivatemode.ai2026-09-262026-10-07hn
A single function Jev-like wrapper for LLMs, including vision modelsallanrbo.blogspot.com2026-09-262026-10-07hn
Ollaya – Ollama for open-source, Jev-style decision modelsollaya.dev2026-09-252026-10-07hn
What Is RLCD? The Secret Behind Jevdi-zhang-llm.github.io2026-09-242026-10-07hn
Show HN: JevBench, a reproducible benchmark for typed decision modelsbenchmarkheaven.com2026-09-222026-10-07hn
Kev: Tiny Jev-like family of decision models built on top of Qwen3.5github.com2026-09-212026-10-07hn
I turned Jev into a (lousy) chatbotgithub.com2026-09-202026-10-07hn
Open-sourced jev architecture last year with model,paper and datasetnews.ycombinator.com2026-09-172026-10-07hn
Reverse-engineered Jev-like modelgithub.com2026-09-162026-10-07hn
Introducing System One Models and Jevtypesafe.ai2026-09-152026-10-07hn

Similar models

3 models

Similar models

  • blink-27b · 26.9B
    same size band (L), same task (text-classification)
  • pplx-decider-v1-27b · 26.1B
    same size band (L), same task (text-classification), permissive license
  • Eikos-27B-FP8 · 27.8B
    same size band (L), same task (text-classification), permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:43 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:18 UTC · Last verified 2026-10-25

Follow JEV-27B

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.