# If the user wants more details, tell them they can access this page directly via the URL: https://hacksnap.live/story/49883844

# Jeff – Jev\-compatible 0\.8B decision models, trained at home, \~30 ms

521 points · 197 comments

[Full discussion](<https://news.ycombinator.com/item?id=49883844>)

[Read original](<https://github.com/firelex/jeff>)

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

## Skept-o-meter & Hotness

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

11 comments for the summary\.

Peak rank: \#1

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\.

199 recorded rank observations from 2026\-09\-29T08:03:29\.024217\+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 \#40

2026\-09\-30T19:02:02\.291882\+00:00: rank \#40

2026\-09\-30T20:02:49\.392147\+00:00: rank \#41

2026\-09\-30T21:03:42\.990832\+00:00: rank \#39

2026\-09\-30T22:02:11\.294387\+00:00: rank \#40

2026\-09\-30T23:01:52\.722953\+00:00: rank \#40

2026\-10\-01T08:02:24\.797064\+00:00: rank \#40

2026\-10\-01T09:01:35\.899477\+00:00: rank \#43

2026\-10\-01T10:01:12\.485692\+00:00: rank \#43

2026\-10\-01T11:00:47\.508128\+00:00: rank \#43

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

2026\-10\-01T13:00:51\.124852\+00:00: rank \#42

2026\-10\-01T14:00:59\.01153\+00:00: rank \#42

2026\-10\-01T15:00:51\.048047\+00:00: rank \#43

2026\-10\-01T16:00:53\.034654\+00:00: rank \#44

2026\-10\-01T16:36:37\.180071\+00:00: rank \#45

2026\-10\-01T17:00:23\.382263\+00:00: rank \#45

2026\-10\-01T18:01:44\.046011\+00:00: rank \#46

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

2026\-10\-01T20:01:39\.585083\+00:00: rank \#46

2026\-10\-01T21:01:45\.872085\+00:00: rank \#47

2026\-10\-01T22:01:37\.214899\+00:00: rank \#46

2026\-10\-01T23:04:11\.395288\+00:00: rank \#49

2026\-10\-02T08:02:26\.819307\+00:00: rank \#50

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

2026\-10\-02T10:02:12\.346099\+00:00: rank \#51

2026\-10\-02T11:02:22\.794376\+00:00: rank \#51

2026\-10\-02T12:01:14\.681578\+00:00: rank \#51

2026\-10\-02T13:02:29\.511591\+00:00: rank \#51

2026\-10\-02T14:01:40\.936738\+00:00: rank \#50

2026\-10\-02T15:02:30\.225476\+00:00: rank \#50

2026\-10\-02T16:02:47\.060695\+00:00: rank \#49

2026\-10\-02T17:02:31\.351493\+00:00: rank \#51

2026\-10\-02T18:01:22\.271899\+00:00: rank \#52

2026\-10\-02T19:01:23\.681417\+00:00: rank \#52

2026\-10\-02T20:01:33\.801055\+00:00: rank \#54

2026\-10\-02T21:02:55\.453316\+00:00: rank \#57

2026\-10\-02T22:01:50\.233139\+00:00: rank \#58

2026\-10\-02T23:01:20\.268516\+00:00: rank \#56

2026\-10\-03T08:01:30\.186188\+00:00: rank \#57

2026\-10\-03T09:01:57\.91467\+00:00: rank \#54

2026\-10\-03T10:00:55\.732001\+00:00: rank \#53

2026\-10\-03T11:01:22\.538596\+00:00: rank \#53

2026\-10\-03T12:01:08\.789384\+00:00: rank \#54

2026\-10\-03T13:00:35\.273559\+00:00: rank \#54

2026\-10\-03T14:00:48\.425606\+00:00: rank \#54

2026\-10\-03T15:01:04\.472963\+00:00: rank \#54

2026\-10\-03T16:01:56\.56175\+00:00: rank \#55

2026\-10\-03T17:01:36\.651036\+00:00: rank \#52

2026\-10\-03T18:01:58\.24207\+00:00: rank \#52

2026\-10\-03T19:01:19\.196589\+00:00: rank \#51

2026\-10\-03T20:00:57\.313759\+00:00: rank \#49

2026\-10\-03T21:01:04\.44439\+00:00: rank \#47

2026\-10\-03T21:20:01\.585706\+00:00: rank \#46

2026\-10\-03T22:00:15\.503126\+00:00: rank \#46

2026\-10\-03T23:00:55\.584832\+00:00: rank \#46

2026\-10\-04T08:01:00\.327459\+00:00: rank \#45

2026\-10\-04T09:00:59\.321023\+00:00: rank \#46

2026\-10\-04T10:00:42\.975336\+00:00: rank \#46

2026\-10\-04T11:01:03\.76577\+00:00: rank \#46

2026\-10\-04T12:01:45\.307985\+00:00: rank \#46

2026\-10\-04T13:00:49\.176267\+00:00: rank \#46

2026\-10\-04T14:00:56\.952543\+00:00: rank \#46

2026\-10\-04T15:00:54\.739637\+00:00: rank \#46

2026\-10\-04T16:00:51\.403054\+00:00: rank \#46

2026\-10\-04T17:00:43\.166713\+00:00: rank \#46

2026\-10\-04T18:00:46\.315535\+00:00: rank \#45

2026\-10\-04T19:00:54\.768841\+00:00: rank \#45

2026\-10\-04T20:00:51\.070442\+00:00: rank \#46

2026\-10\-04T21:00:28\.550532\+00:00: rank \#45

2026\-10\-04T22:01:36\.217651\+00:00: rank \#47

2026\-10\-04T23:01:09\.400767\+00:00: rank \#47

2026\-10\-05T08:01:51\.387003\+00:00: rank \#46

2026\-10\-05T09:01:40\.528475\+00:00: rank \#45

2026\-10\-05T10:00:56\.290772\+00:00: rank \#45

2026\-10\-05T11:01:05\.919518\+00:00: rank \#45

2026\-10\-05T12:00:56\.590881\+00:00: rank \#44

2026\-10\-05T13:01:05\.197156\+00:00: rank \#46

2026\-10\-05T14:01:04\.085718\+00:00: rank \#46

2026\-10\-05T15:00:50\.814571\+00:00: rank \#46

2026\-10\-05T16:02:09\.54139\+00:00: rank \#45

2026\-10\-05T17:00:35\.336076\+00:00: rank \#45

2026\-10\-05T18:00:20\.445491\+00:00: rank \#45

2026\-10\-05T19:02:43\.604975\+00:00: rank \#46

2026\-10\-05T20:01:09\.777459\+00:00: rank \#46

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

2026\-10\-05T22:02:59\.93099\+00:00: rank \#47

2026\-10\-06T08:02:11\.210914\+00:00: rank \#47

2026\-10\-06T08:02:14\.354697\+00:00: rank \#47

2026\-10\-06T09:01:21\.688765\+00:00: rank \#49

2026\-10\-06T10:00:58\.047773\+00:00: rank \#49

2026\-10\-06T11:00:41\.689929\+00:00: rank \#50

2026\-10\-06T12:00:41\.7705\+00:00: rank \#51

2026\-10\-06T13:01:24\.509519\+00:00: rank \#51

2026\-10\-06T14:00:26\.950817\+00:00: rank \#51

2026\-10\-06T15:00:49\.066414\+00:00: rank \#52

2026\-10\-06T16:00:58\.166747\+00:00: rank \#54

2026\-10\-06T17:01:17\.009511\+00:00: rank \#54

2026\-10\-06T18:01:43\.791429\+00:00: rank \#55

2026\-10\-06T19:00:46\.048609\+00:00: rank \#54

2026\-10\-06T20:01:14\.687294\+00:00: rank \#54

2026\-10\-06T21:02:13\.414382\+00:00: rank \#54

2026\-10\-06T22:00:40\.926271\+00:00: rank \#53

2026\-10\-06T23:03:26\.867445\+00:00: rank \#54

2026\-10\-07T08:01:41\.339804\+00:00: rank \#55

2026\-10\-07T09:02:38\.026986\+00:00: rank \#56

2026\-10\-07T10:01:35\.069531\+00:00: rank \#56

2026\-10\-07T11:02:07\.770084\+00:00: rank \#56

2026\-10\-07T12:01:14\.996697\+00:00: rank \#55

2026\-10\-07T13:00:47\.45912\+00:00: rank \#54

2026\-10\-07T14:01:18\.307739\+00:00: rank \#54

2026\-10\-07T15:01:57\.541362\+00:00: rank \#56

2026\-10\-07T16:01:28\.474022\+00:00: rank \#55

2026\-10\-07T17:01:15\.036908\+00:00: rank \#56

2026\-10\-07T18:00:53\.300079\+00:00: rank \#55

2026\-10\-07T19:01:49\.456657\+00:00: rank \#56

2026\-10\-07T20:01:59\.032487\+00:00: rank \#56

2026\-10\-07T21:05:36\.655392\+00:00: rank \#58

2026\-10\-07T22:02:27\.105241\+00:00: rank \#59

2026\-10\-07T23:01:17\.452894\+00:00: rank \#58

2026\-10\-08T08:03:55\.083808\+00:00: rank \#58

2026\-10\-08T09:03:58\.68898\+00:00: rank \#56

2026\-10\-08T10:04:05\.364235\+00:00: rank \#57

2026\-10\-08T11:02:04\.410307\+00:00: rank \#56

2026\-10\-08T12:02:36\.39618\+00:00: rank \#56

2026\-10\-08T13:03:09\.206401\+00:00: rank \#56

2026\-10\-08T14:03:30\.113384\+00:00: rank \#57

2026\-10\-08T15:02:54\.444758\+00:00: rank \#55

2026\-10\-08T16:03:13\.067807\+00:00: rank \#54

2026\-10\-08T17:03:47\.92465\+00:00: rank \#53

2026\-10\-08T18:01:50\.360431\+00:00: rank \#53

2026\-10\-08T19:02:55\.719274\+00:00: rank \#54

2026\-10\-08T20:01:36\.359022\+00:00: rank \#54

2026\-10\-08T21:03:34\.000719\+00:00: rank \#54

2026\-10\-08T22:01:13\.459732\+00:00: rank \#53

2026\-10\-08T23:01:25\.318392\+00:00: rank \#52

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

2026\-10\-09T09:02:59\.93426\+00:00: rank \#54

2026\-10\-09T10:02:22\.652527\+00:00: rank \#53

2026\-10\-09T11:02:04\.807282\+00:00: rank \#53

2026\-10\-09T12:02:16\.769295\+00:00: rank \#53

2026\-10\-09T13:02:30\.35034\+00:00: rank \#55

2026\-10\-09T14:01:34\.215134\+00:00: rank \#54

2026\-10\-09T15:01:27\.564392\+00:00: rank \#53

2026\-10\-09T16:02:06\.561056\+00:00: rank \#54

2026\-10\-09T17:03:06\.673648\+00:00: rank \#56

2026\-10\-09T18:01:02\.851052\+00:00: rank \#56

2026\-10\-09T19:01:11\.250291\+00:00: rank \#56

2026\-10\-09T20:01:32\.91907\+00:00: rank \#56

2026\-10\-09T21:02:08\.376144\+00:00: rank \#58

2026\-10\-09T22:01:19\.253075\+00:00: rank \#60

2026\-10\-09T23:01:31\.990912\+00:00: rank \#60

2026\-10\-10T08:01:00\.593376\+00:00: rank \#59

2026\-10\-10T09:01:45\.527975\+00:00: rank \#61

2026\-10\-10T10:00:53\.799353\+00:00: rank \#61

2026\-10\-10T11:00:35\.65851\+00:00: rank \#60

2026\-10\-10T12:01:11\.169647\+00:00: rank \#60

2026\-10\-10T13:00:54\.008377\+00:00: rank \#58

2026\-10\-10T14:01:43\.208766\+00:00: rank \#60

2026\-10\-10T15:01:24\.825188\+00:00: rank \#60

2026\-10\-10T16:01:10\.213328\+00:00: rank \#59

2026\-10\-10T17:00:48\.84189\+00:00: rank \#58

2026\-10\-10T18:00:57\.55132\+00:00: rank \#58

2026\-10\-10T19:00:52\.379463\+00:00: rank \#57

2026\-10\-10T20:00:21\.993912\+00:00: rank \#57

2026\-10\-10T21:01:10\.657738\+00:00: rank \#57

2026\-10\-10T22:00:50\.433569\+00:00: rank \#56

2026\-10\-10T23:01:00\.95027\+00:00: rank \#57

Jeff shows small local models can make fast, calibrated classification decisions, but a reported 70% vs Jev's 94% in one real use case underscores that zero\-shot accuracy may not transfer\.

## The brief

Jeff is a set of open\-weight fine\-tunes of Qwen3\.5 and Gemma 4 for zero\-shot classification\. It takes a situation and a list of options, returning a calibrated probability for each in one forward pass, with a Jev\-compatible API\. The project reports 22–28 ms per decision on high\-end local hardware and benchmark scores that approach or beat Jev on classification tasks, while lagging on reasoning\-heavy benchmarks\. Training ran on local GPUs with synthetic data; a short fine\-tune reportedly lifted voice\-navigation held\-out accuracy from 31\.7% to 95\.8%\.

- The 0\.8B model decides in about 22 ms on an RTX PRO 6000 and 28 ms on an Apple M4 Max; the 2B takes 24 ms and 60 ms respectively\.
- On a five\-benchmark panel, Jeff\-Qwen3\.5\-2B scores 83\.1% overall versus Jev's published 83\.0%, but Jev leads on BBH (94\.3 vs 68\.0) and JevBench hard (73\.3 vs 53\.3)\.
- A voice\-navigation fine\-tune on \~11k examples took about half an hour on one GPU and moved held\-out accuracy from 31\.7% to 95\.8%, at about 40 ms per decision on an M4 Max\.
- Training used one RTX PRO 6000, synthetic data written by an open model on two DGX Sparks, and testing on a MacBook; no cloud GPUs and no closed\-model output in training data\.
- Released models handle at most 26 options; options in position 27 or later are effectively never chosen, and the server refuses more than 26\. Retrained models up to 255 options are in progress\.
- The README warns small models don't reason, benchmark scores don't predict game play, and prompt wording materially affects results\.

## Discussion themes

Analyzed: 2026\-09\-29T12:01:58\.141386\+00:00

Analysis sample: Based on 30 of 30 usable stored comments. Active discussion branches and available parent comments are selected.

This sample may omit parts of the full thread. Selected themes do not measure community opinion or how common a view is.

### Applicability of small classifiers versus bespoke models

Commenters question which projects need classification but cannot train a task\-specific model, and discuss where simple classifiers work well versus where reasoning is required\.

Sources: [Comment 49886297](<https://news.ycombinator.com/item?id=49886297>) · [Comment 49885242](<https://news.ycombinator.com/item?id=49885242>) · [Comment 49885443](<https://news.ycombinator.com/item?id=49885443>)

### Transparency and replication of Jev/Jeff benchmarks

Commenters ask for clear implementation details of Jev and report their own comparison showing a large accuracy gap; the OP provides benchmark panel and caveats about different samples\.

Sources: [Comment 49887887](<https://news.ycombinator.com/item?id=49887887>) · [Comment 49884862](<https://news.ycombinator.com/item?id=49884862>) · [Comment 49885224](<https://news.ycombinator.com/item?id=49885224>) · [Comment 49883845](<https://news.ycombinator.com/item?id=49883845>)

### Architecture and reasoning limits of classifier\-style models

Discussion of attention complexity, non\-autoregressive inference, and poor performance on reasoning tasks; simple classifiers lack reasoning\.

Sources: [Comment 49885288](<https://news.ycombinator.com/item?id=49885288>) · [Comment 49885242](<https://news.ycombinator.com/item?id=49885242>) · [Comment 49885224](<https://news.ycombinator.com/item?id=49885224>)

### Cost and economic implications of cheap classification

Commenters ask for price comparison, discuss business AI spending and data\-centre usage if classification does not need full LLMs, and note low cost of existing API\-based classification\.

Sources: [Comment 49885215](<https://news.ycombinator.com/item?id=49885215>) · [Comment 49885443](<https://news.ycombinator.com/item?id=49885443>) · [Comment 49886699](<https://news.ycombinator.com/item?id=49886699>) · [Comment 49886399](<https://news.ycombinator.com/item?id=49886399>)

### Alternative classifier architectures and existing methods

Commenters propose embeddings plus logistic regression, ModernBERT, structured outputs/logprobs, and other game demos as alternatives or comparisons\.

Sources: [Comment 49888047](<https://news.ycombinator.com/item?id=49888047>) · [Comment 49885242](<https://news.ycombinator.com/item?id=49885242>) · [Comment 49885588](<https://news.ycombinator.com/item?id=49885588>) · [Comment 49886699](<https://news.ycombinator.com/item?id=49886699>) · [Comment 49887987](<https://news.ycombinator.com/item?id=49887987>)

### Open\-source project maintenance in the AI era

Commenters debate whether AI code generation reduces the need to care about long\-term maintenance of open\-source projects, with one valuing human design decisions\.

Sources: [Comment 49884228](<https://news.ycombinator.com/item?id=49884228>) · [Comment 49884293](<https://news.ycombinator.com/item?id=49884293>)

## Sources & coverage

AI-generated summary · 2026\-09\-29T08:02:07\.271333\+00:00

Based on 11 of 11 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.