# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/49783999

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

390 points · 172 comments

[Full discussion](<https://news.ycombinator.com/item?id=49783999>)

[Read original](<https://github.com/jaredpalmer/kev/tree/main>)

Category: [Models & Products](<https://hacksnap.live/?category=models-products>)

## Skept-o-meter & Hotness

Skept\-o\-meter: Low\. Estimated from 10 comments\.

7 comments for the summary\.

Peak rank: \#3

Time in Top 10: 24\.0 hours

Hacksnap ranks recent stories first, then orders each group by points\. Peak rank uses all retained observations\. Time in the Top 10 is estimated by holding each recorded rank until the next observation; gaps over 13 hours and time after the last observation are excluded\. Movement between observations is unknown\.

237 recorded rank observations from 2026\-09\-21T12:01:17\.894498\+00:00 to 2026\-10\-10T23:01:00\.95027\+00:00\.

Hotness — latest 168 recorded Hacksnap ranks:

2026\-09\-30T18:03:34\.83179\+00:00: rank \#57

2026\-09\-30T19:02:02\.291882\+00:00: rank \#57

2026\-09\-30T20:02:49\.392147\+00:00: rank \#58

2026\-09\-30T21:03:42\.990832\+00:00: rank \#56

2026\-09\-30T22:02:11\.294387\+00:00: rank \#57

2026\-09\-30T23:01:52\.722953\+00:00: rank \#57

2026\-10\-01T08:02:24\.797064\+00:00: rank \#57

2026\-10\-01T09:01:35\.899477\+00:00: rank \#60

2026\-10\-01T10:01:12\.485692\+00:00: rank \#60

2026\-10\-01T11:00:47\.508128\+00:00: rank \#60

2026\-10\-01T12:00:42\.477917\+00:00: rank \#59

2026\-10\-01T13:00:51\.124852\+00:00: rank \#59

2026\-10\-01T14:00:59\.01153\+00:00: rank \#59

2026\-10\-01T15:00:51\.048047\+00:00: rank \#60

2026\-10\-01T16:00:53\.034654\+00:00: rank \#61

2026\-10\-01T16:36:37\.180071\+00:00: rank \#62

2026\-10\-01T17:00:23\.382263\+00:00: rank \#62

2026\-10\-01T18:01:44\.046011\+00:00: rank \#63

2026\-10\-01T19:01:47\.420738\+00:00: rank \#64

2026\-10\-01T20:01:39\.585083\+00:00: rank \#63

2026\-10\-01T21:01:45\.872085\+00:00: rank \#64

2026\-10\-01T22:01:37\.214899\+00:00: rank \#63

2026\-10\-01T23:04:11\.395288\+00:00: rank \#66

2026\-10\-02T08:02:26\.819307\+00:00: rank \#67

2026\-10\-02T09:02:51\.51649\+00:00: rank \#68

2026\-10\-02T10:02:12\.346099\+00:00: rank \#68

2026\-10\-02T11:02:22\.794376\+00:00: rank \#68

2026\-10\-02T12:01:14\.681578\+00:00: rank \#68

2026\-10\-02T13:02:29\.511591\+00:00: rank \#68

2026\-10\-02T14:01:40\.936738\+00:00: rank \#67

2026\-10\-02T15:02:30\.225476\+00:00: rank \#67

2026\-10\-02T16:02:47\.060695\+00:00: rank \#66

2026\-10\-02T17:02:31\.351493\+00:00: rank \#68

2026\-10\-02T18:01:22\.271899\+00:00: rank \#69

2026\-10\-02T19:01:23\.681417\+00:00: rank \#69

2026\-10\-02T20:01:33\.801055\+00:00: rank \#71

2026\-10\-02T21:02:55\.453316\+00:00: rank \#74

2026\-10\-02T22:01:50\.233139\+00:00: rank \#75

2026\-10\-02T23:01:20\.268516\+00:00: rank \#73

2026\-10\-03T08:01:30\.186188\+00:00: rank \#74

2026\-10\-03T09:01:57\.91467\+00:00: rank \#73

2026\-10\-03T10:00:55\.732001\+00:00: rank \#72

2026\-10\-03T11:01:22\.538596\+00:00: rank \#72

2026\-10\-03T12:01:08\.789384\+00:00: rank \#73

2026\-10\-03T13:00:35\.273559\+00:00: rank \#73

2026\-10\-03T14:00:48\.425606\+00:00: rank \#73

2026\-10\-03T15:01:04\.472963\+00:00: rank \#73

2026\-10\-03T16:01:56\.56175\+00:00: rank \#74

2026\-10\-03T17:01:36\.651036\+00:00: rank \#71

2026\-10\-03T18:01:58\.24207\+00:00: rank \#71

2026\-10\-03T19:01:19\.196589\+00:00: rank \#70

2026\-10\-03T20:00:57\.313759\+00:00: rank \#69

2026\-10\-03T21:01:04\.44439\+00:00: rank \#67

2026\-10\-03T21:20:01\.585706\+00:00: rank \#66

2026\-10\-03T22:00:15\.503126\+00:00: rank \#66

2026\-10\-03T23:00:55\.584832\+00:00: rank \#66

2026\-10\-04T08:01:00\.327459\+00:00: rank \#65

2026\-10\-04T09:00:59\.321023\+00:00: rank \#66

2026\-10\-04T10:00:42\.975336\+00:00: rank \#66

2026\-10\-04T11:01:03\.76577\+00:00: rank \#66

2026\-10\-04T12:01:45\.307985\+00:00: rank \#66

2026\-10\-04T13:00:49\.176267\+00:00: rank \#66

2026\-10\-04T14:00:56\.952543\+00:00: rank \#66

2026\-10\-04T15:00:54\.739637\+00:00: rank \#66

2026\-10\-04T16:00:51\.403054\+00:00: rank \#67

2026\-10\-04T17:00:43\.166713\+00:00: rank \#67

2026\-10\-04T18:00:46\.315535\+00:00: rank \#66

2026\-10\-04T19:00:54\.768841\+00:00: rank \#66

2026\-10\-04T20:00:51\.070442\+00:00: rank \#67

2026\-10\-04T21:00:28\.550532\+00:00: rank \#66

2026\-10\-04T22:01:36\.217651\+00:00: rank \#68

2026\-10\-04T23:01:09\.400767\+00:00: rank \#68

2026\-10\-05T08:01:51\.387003\+00:00: rank \#67

2026\-10\-05T09:01:40\.528475\+00:00: rank \#66

2026\-10\-05T10:00:56\.290772\+00:00: rank \#66

2026\-10\-05T11:01:05\.919518\+00:00: rank \#66

2026\-10\-05T12:00:56\.590881\+00:00: rank \#65

2026\-10\-05T13:01:05\.197156\+00:00: rank \#67

2026\-10\-05T14:01:04\.085718\+00:00: rank \#67

2026\-10\-05T15:00:50\.814571\+00:00: rank \#67

2026\-10\-05T16:02:09\.54139\+00:00: rank \#66

2026\-10\-05T17:00:35\.336076\+00:00: rank \#66

2026\-10\-05T18:00:20\.445491\+00:00: rank \#66

2026\-10\-05T19:02:43\.604975\+00:00: rank \#67

2026\-10\-05T20:01:09\.777459\+00:00: rank \#67

2026\-10\-05T21:01:47\.650671\+00:00: rank \#68

2026\-10\-05T22:02:59\.93099\+00:00: rank \#69

2026\-10\-06T08:02:11\.210914\+00:00: rank \#69

2026\-10\-06T08:02:14\.354697\+00:00: rank \#69

2026\-10\-06T09:01:21\.688765\+00:00: rank \#71

2026\-10\-06T10:00:58\.047773\+00:00: rank \#71

2026\-10\-06T11:00:41\.689929\+00:00: rank \#72

2026\-10\-06T12:00:41\.7705\+00:00: rank \#73

2026\-10\-06T13:01:24\.509519\+00:00: rank \#73

2026\-10\-06T14:00:26\.950817\+00:00: rank \#73

2026\-10\-06T15:00:49\.066414\+00:00: rank \#74

2026\-10\-06T16:00:58\.166747\+00:00: rank \#76

2026\-10\-06T17:01:17\.009511\+00:00: rank \#76

2026\-10\-06T18:01:43\.791429\+00:00: rank \#77

2026\-10\-06T19:00:46\.048609\+00:00: rank \#77

2026\-10\-06T20:01:14\.687294\+00:00: rank \#77

2026\-10\-06T21:02:13\.414382\+00:00: rank \#77

2026\-10\-06T22:00:40\.926271\+00:00: rank \#76

2026\-10\-06T23:03:26\.867445\+00:00: rank \#78

2026\-10\-07T08:01:41\.339804\+00:00: rank \#79

2026\-10\-07T09:02:38\.026986\+00:00: rank \#80

2026\-10\-07T10:01:35\.069531\+00:00: rank \#80

2026\-10\-07T11:02:07\.770084\+00:00: rank \#80

2026\-10\-07T12:01:14\.996697\+00:00: rank \#79

2026\-10\-07T13:00:47\.45912\+00:00: rank \#78

2026\-10\-07T14:01:18\.307739\+00:00: rank \#78

2026\-10\-07T15:01:57\.541362\+00:00: rank \#80

2026\-10\-07T16:01:28\.474022\+00:00: rank \#80

2026\-10\-07T17:01:15\.036908\+00:00: rank \#81

2026\-10\-07T18:00:53\.300079\+00:00: rank \#80

2026\-10\-07T19:01:49\.456657\+00:00: rank \#81

2026\-10\-07T20:01:59\.032487\+00:00: rank \#81

2026\-10\-07T21:05:36\.655392\+00:00: rank \#83

2026\-10\-07T22:02:27\.105241\+00:00: rank \#84

2026\-10\-07T23:01:17\.452894\+00:00: rank \#83

2026\-10\-08T08:03:55\.083808\+00:00: rank \#83

2026\-10\-08T09:03:58\.68898\+00:00: rank \#81

2026\-10\-08T10:04:05\.364235\+00:00: rank \#82

2026\-10\-08T11:02:04\.410307\+00:00: rank \#81

2026\-10\-08T12:02:36\.39618\+00:00: rank \#81

2026\-10\-08T13:03:09\.206401\+00:00: rank \#81

2026\-10\-08T14:03:30\.113384\+00:00: rank \#82

2026\-10\-08T15:02:54\.444758\+00:00: rank \#80

2026\-10\-08T16:03:13\.067807\+00:00: rank \#79

2026\-10\-08T17:03:47\.92465\+00:00: rank \#78

2026\-10\-08T18:01:50\.360431\+00:00: rank \#78

2026\-10\-08T19:02:55\.719274\+00:00: rank \#79

2026\-10\-08T20:01:36\.359022\+00:00: rank \#79

2026\-10\-08T21:03:34\.000719\+00:00: rank \#79

2026\-10\-08T22:01:13\.459732\+00:00: rank \#78

2026\-10\-08T23:01:25\.318392\+00:00: rank \#77

2026\-10\-09T08:01:53\.314021\+00:00: rank \#78

2026\-10\-09T09:02:59\.93426\+00:00: rank \#79

2026\-10\-09T10:02:22\.652527\+00:00: rank \#78

2026\-10\-09T11:02:04\.807282\+00:00: rank \#78

2026\-10\-09T12:02:16\.769295\+00:00: rank \#78

2026\-10\-09T13:02:30\.35034\+00:00: rank \#80

2026\-10\-09T14:01:34\.215134\+00:00: rank \#79

2026\-10\-09T15:01:27\.564392\+00:00: rank \#78

2026\-10\-09T16:02:06\.561056\+00:00: rank \#79

2026\-10\-09T17:03:06\.673648\+00:00: rank \#81

2026\-10\-09T18:01:02\.851052\+00:00: rank \#81

2026\-10\-09T19:01:11\.250291\+00:00: rank \#81

2026\-10\-09T20:01:32\.91907\+00:00: rank \#81

2026\-10\-09T21:02:08\.376144\+00:00: rank \#83

2026\-10\-09T22:01:19\.253075\+00:00: rank \#85

2026\-10\-09T23:01:31\.990912\+00:00: rank \#85

2026\-10\-10T08:01:00\.593376\+00:00: rank \#84

2026\-10\-10T09:01:45\.527975\+00:00: rank \#86

2026\-10\-10T10:00:53\.799353\+00:00: rank \#86

2026\-10\-10T11:00:35\.65851\+00:00: rank \#85

2026\-10\-10T12:01:11\.169647\+00:00: rank \#85

2026\-10\-10T13:00:54\.008377\+00:00: rank \#83

2026\-10\-10T14:01:43\.208766\+00:00: rank \#85

2026\-10\-10T15:01:24\.825188\+00:00: rank \#85

2026\-10\-10T16:01:10\.213328\+00:00: rank \#84

2026\-10\-10T17:00:48\.84189\+00:00: rank \#83

2026\-10\-10T18:00:57\.55132\+00:00: rank \#83

2026\-10\-10T19:00:52\.379463\+00:00: rank \#82

2026\-10\-10T20:00:21\.993912\+00:00: rank \#82

2026\-10\-10T21:01:10\.657738\+00:00: rank \#83

2026\-10\-10T22:00:50\.433569\+00:00: rank \#82

2026\-10\-10T23:01:00\.95027\+00:00: rank \#83

Kev is a practical, transparently benchmarked local decision\-model stack, but the discussion suggests its novelty and 'Jev\-like' label are still contested, with open questions about use cases and the architectural or training differences from Jev\.

## The brief

Kev is an Apache\-2\.0 family of small decision models (0\.8B, 4B, and 9B) built on Qwen3\.5 bases and inspired by Jev's architecture\. It provides LoRA adapters plus a pointer head, a local server/API compatible with TypeSafe's System One, noul/choice/score question types, training and evaluation tooling, and benchmarks against Jev and the prior Qwen3 generation\. The README reports strong in\-distribution accuracy, weaker new\-source and knowledge performance, calibration improvements via optional temperature and date\-fact settings, and limitations including option\-order sensitivity and slow Apple Silicon serving\.

- Kev\-0\.8B, Kev\-4B, and Kev\-9B are built on Qwen3\.5 bases with rank\-16 LoRA adapters and a small pointer head; the earlier Qwen3\-based generation remains published but is no longer developed\.
- The local API exposes POST /v1/systemone with noul, choice, and score questions in one request, and is designed to work with TypeSafe's Python SDK; the server binds to localhost and has no authentication\.
- On attention\-only Qwen3 bases, state and questions share one masked sequence; on Qwen3\.5, Gated DeltaNet layers require each question to run as its own row, with the state cache reused across rows\.
- Reported accuracy ranges from 0\.825 to 0\.874 on trained sources and 0\.652 to 0\.852 on new sources; Kev\-9B trails Jev on the new\-source development set (0\.822 vs 0\.857), though the README says this is not a controlled architecture comparison\.
- Training uses decision\-v7 with 10,000 examples from ten public datasets plus generated policy and rule examples; fine\-tuning with \-\-init\_from is recommended over starting from the base model, which in one user test scored 0\.33 versus 0\.84 on Kev's evaluation set\.
- Limitations include overconfidence on new sources, date\-arithmetic failures, a large MMLU/MMLU\-Pro gap versus Jev, option\-order sensitivity, slow Apple Silicon serving for Qwen3\.5 models, and no request batching or authentication\.

## Discussion themes

### Use cases remain unclear

One commenter is impressed by the open\-source energy around Jev\-like models but asks what these models are actually for, including whether they fit coding agents or entirely different situations\.

Sources: [Comment 49784255](<https://news.ycombinator.com/item?id=49784255>)

### A 'Jev explosion' and the value of developer communication

Commenters frame the trend as a Jev explosion and note that classification models have existed for a while\. A reply argues the real contribution is presenting ML to ordinary developers through a simple JSON\-style interface, similar to what Ansible did for configuration management\.

Sources: [Comment 49784304](<https://news.ycombinator.com/item?id=49784304>) · [Comment 49784835](<https://news.ycombinator.com/item?id=49784835>)

### Clarifying the classification model

A commenter asks how Jev differs from a normal neural network classifier, describing it as text input plus one\-shot classification without training data\. A reply confirms this is essentially right and adds that the output classes can be arbitrary rather than fixed before training\.

Sources: [Comment 49784613](<https://news.ycombinator.com/item?id=49784613>) · [Comment 49784632](<https://news.ycombinator.com/item?id=49784632>)

### Skepticism about the 'Jev\-like' label

One commenter questions whether a model built on a Qwen base trained with RLHF can be considered Jev\-like if Jev is fundamentally trained using RLCD\. Another commenter responds with an anecdote that Jev, when given alphabet choices and asked in a loop what model it is, reportedly says Qwen\.

Sources: [Comment 49784983](<https://news.ycombinator.com/item?id=49784983>) · [Comment 49785351](<https://news.ycombinator.com/item?id=49785351>)

## Sources & coverage

AI-generated summary · 2026\-09\-21T12:01:15\.117984\+00:00

Based on 7 of 7 usable stored comments, selected by depth and branch activity. This is a sample of the discussion. Article text may also be shortened.

Generated using deepseek\-ai/DeepSeek\-V4\.1\-Flash. Check the linked sources for full context.