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 › kev family

kev-9b

Hugging Face ↗ Follow changes

jaredpalmer Apache 2.0 out 35 days ago

Text in, labels out.

Verified 2026-10-25 Rank #0 in Decision models · score 0 method v1 Sources 3 Downloads 30d 6.2k · #530 of 619 tracked

Can I use it?Apache 2.0Permissive; patent grantlicence file2026-10-24
Do people use it?6.2kdownloads, 30 days#530 of 619 we track · below the median+3 likes, downloads −11% this week81 likesHugging Face2026-10-24

Worth reading

Picked by rule from 1 tracked piece: the lab's own post, the most shared, a field report, the latest.

Most discussed · 463 HN pts

Kev: Tiny Jev-like family of decision models built on top of Qwen3.5

Kev is a family of small decision models built on Qwen3.5 and Qwen3.8 that you can train and run locally.

  • 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.
github.com 2026-09-21 463 HN pts · 210 comments 6.7k words about the kev family family

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 9 Spaces.

  • As a service0no API provider listed
  • On own hardware2quantized builds
  • Built on31 fine-tunes · 1 adapters · 1 merges
  • In apps9Hugging Face Spaces

Top Spaces running it: multimodalart/jev-decision-index ♥ 802

The story so far

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

1 dated article and 8 repo commits so far

Teal is the weights (2026-09-20), amber triangles are 1 dated article (each links to it), grey ticks are repo commits.

Oct2026-10-01: 1 repo commit2026-09-30: 1 repo commit2026-09-21: 4 repo commits2026-09-20: 2 repo commitsKev: Tiny Jev-like family of decision models built on top of Qwen3.5 · 2026-09-21Weights 2026-09-20 · 2026-09-20Weights 2026-09-20now
Oct2026-10-01: 1 repo commit2026-09-30: 1 repo commit2026-09-21: 4 repo commits2026-09-20: 2 repo commitsKev: Tiny Jev-like family of decision models built on top of Qwen3.5 · 2026-09-21Weights 2026-09-20 · 2026-09-20Weights 2026-09-20now

What people built on it

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

5 derivatives on the Hub, led by 2 quantized

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

2 quantized (Hugging Face model tree)1 fine-tune (Hugging Face model tree)1 merge (Hugging Face model tree)1 adapter (Hugging Face model tree)2 quantized1 fine-tune1 merge1 adapter
2 quantized (Hugging Face model tree)1 fine-tune (Hugging Face model tree)1 merge (Hugging Face model tree)1 adapter (Hugging Face model tree)2 quantized1 fine-tune1 merge1 adapter
GGUF ggml-org/Kev-9B-GGUF for llama.cpp, Ollama, LM Studio 5.8k 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 MMLU-Pro · Reason through hard problems0.59—model author2026-10-07
author-reported Out-of-domain, transfer-v4 development (656): accuracy / Brier0.820 / 0.262—model author2026-10-07
author-reported Out-of-domain, transfer-v4 locked test (656): accuracy0.85—model author2026-10-07
author-reported Short states, transfer-r3 test (1,150)0.85—model author2026-10-07
author-reported Short states: transfer-v4 development + transfer-r3 test, without `emotion` (1,586)0.87—model author2026-10-07
author-reported Unanswerable items answered with p ≥ 0.9 (lower is better)0—model author2026-10-07
author-reported breadth-v1 development, all 14 datasets (3,075)0.7—model author2026-10-07
author-reported breadth-v1 test, all 14 datasets (3,089)0.7—model author2026-10-07
author-reported breadth-v1 test, chance-corrected index¹ [95 % CI]41—model author2026-10-07
author-reported documents-v1 development (920)0.9—model author2026-10-07
author-reported documents-v1 test (936)0.9—model author2026-10-07
author-reported hard-v1 + devtools-v1 development, audited (1,855)0.82—model author2026-10-07
author-reported hard-v1 + devtools-v1 test, audited (1,859)0.82—model author2026-10-07
author-reported transfer-v4 locked test: ECE / confident errors (p ≥ 0.9 and wrong) / coverage at ≤ 5 % error0.034 / 1.4% / 0.742—model author2026-10-07
author-reported transfer-v4 locked test: served Brier0.2—model author2026-10-07

What the card says

3 cited sentences

Summary generated, cited

Kev-9B is an English decision model that functions as a LoRA adapter and pointer head on the Qwen3.5-9B-Base backbone.[1] The model is designed for document classification, routing, and triage by returning calibrated probability distributions without generating text.[2] It is released under the Apache 2.0 license and can fit on a single 24 GB-class GPU.[3][4]

  1. “LoRA adapter and a pointer head on Qwen3.5-9B-Base”
  2. “returns a calibrated probability distribution over the op…”
  3. license
  4. “fits one 24 GB-class GPU”

Generated 2026-10-07 15:18 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

5 use cases

What it is good at 5 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

12 fields

Listed in Decision models

Facts

Pipeline tagtext-classification
Librarypeft
MoEno
Languagesen
LicenseApache 2.0 (apache-2.0)
Hub snapshot2026-10-25

Decision facts

catalog

Decision-model catalog curated

How it was madeLoRA + new scoring/pointer head
Lineage bucketF - fine-tune of an open model (public/synthetic data)
Lineage evidenceboard file inspection (adapter/head/weight files) + HF base_model field
Base modelQwen/Qwen3.5-9B-Base
DI 0.2.1 score38.48 (independent (DI 0.2.1 board))
ECE (board)0.1378
Median latency51.4 ms (DI board, 1x RTX PRO 6000, single process)
Hosted price (input)self-hosted/M
Sourcehttps://huggingface.co/jaredpalmer/kev-9b

All variants

5 variants

Lineage

Adapter Family: jaredpalmer/kev 5 derivatives on the Hub

Base model: Qwen/Qwen3.5-9B-Base

quantized: 2adapter: 1merge: 1finetune: 1

Top variants (5 shown, ranked by downloads)

Relation
5 rows
Variant Relation Quant Params Band License Author Likes Downloads
ggml-org/Kev-9B-GGUF quantized GGUF — — apache-2.0 ggml-org 4 5.8k
turintech/Kev-9B-Artemis-GGUF quantized GGUF — — apache-2.0 turintech 8 1.3k
gyung/Qwev-9B-RLCD adapter — — — apache-2.0 gyung 2 92
ayan4m1/kev-9b-merged merge — 9.65B L apache-2.0 ayan4m1 0 122
AIMultiple/kev-9b-browser-ft finetune — — — apache-2.0 AIMultiple 0 38

Repository changelog

8 commits

Repository changelog 8 commits

DateCommitSHA
2026-10-01 Kev 1.0 model card (weights unchanged) db029f0
2026-09-30 Kev-9B v2: round 27's confirmed candidate (round 18's documents + skills delta on v1), T 2.19; v1 at tag v1 b5d8c18
2026-09-21 card wording 2629c06
2026-09-21 calibration built in: head.pt carries the fitted temperature; KEV_TEMPERATURE=1.0 for raw logits 3cf1ab7
2026-09-21 Model card: link the Space demo b545836
2026-09-21 Kev-9B: dates+unknowable delta on the v7 checkpoint (locked test 0.874 / 0.852); previous weights at tag v7-base 442e597
2026-09-20 Kev-9B: Qwen3.5-9B-Base, decision-v7 recipe (locked test 0.873 / 0.837) 6281032
2026-09-20 initial commit 5e14563

Articles and coverage

1 piece

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.

Analysis1

Kev: Tiny Jev-like family of decision models built on top of Qwen3.5
github.com2026-09-21authorityHN front pageabout the kev 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.

Similar models

3 models

Similar models

  • Bespoke-Nimble-9B
    same task (text-classification), permissive license
  • Solomon
    same task (text-classification), permissive license
  • NeoHorse-Jev-4B
    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 kev-9b

Watch this model

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