The Best Decision Model, Overall: Strands Decider v22, One 4B Model for Text and Images, #1 on the JevBench Text Leaderboard (est.)

By my benchmarks, strands-decider-4B-hobson-v22 (unofficial) is the best decision model on the JevBench text leaderboard. Its estimated composite of about 74 puts it ahead of all 109 ranked systems, including Cygnet (73.70, #1 today), Winnow-12B (73.23, #2) and Jev 1.13.0 (72.13, #3), and its estimated Capability of about 75 puts it first in the 4B class, ahead of jevk5-v0.3-4b (72.3). It does this as one 4B checkpoint that also answers questions about images. These positions are local, unofficial estimates until the official run measures them. v22 is a community release built on Strands Decider and proposed to the Strands team; it is not an official Strands release.

strands-decider-4B-hobson-v22 is a single 4B Strands Decider checkpoint for text and images. It scores 64.6 v1.5-style Intelligence on the 231 public JevBench v1 tasks, against 53.6 for my best 2B model from the previous article and 29.5 for the official Strands Decider v19. On held-out task families it scores 24.7, against 8.4 for my best 2B, and on images it leads every published 2B model I measured.

Alt text: Bar charts of public text score (v19 29.5, best 2B 53.6, 4B pilot 50.4, first 4B run 56.1, v22 64.6) and unseen-v2 test score (v19 -4.1, best 2B 6.3, first 4B run 15.3, v22 24.7).

My estimate places v22 #1 overall on the JevBench text board, with a composite of about 74 against 73.70 for Cygnet (#1 today), 73.23 for Winnow-12B (#2) and 72.13 for Jev 1.13.0 (#3), and #1 in the 4B class, with Capability about 75 against 72.3 for jevk5-v0.3-4b. The margin over Cygnet is inside the board's tie band, so the official run decides it. These ranks are local, unofficial estimates pending an official run, which alone measures the sealed half and the board's Calibration.

The previous article ended with four 2B-class candidates, offered to the maintainers in issue #24 . The 2B class tops out well below the overall leaders, and every ranked 4B decider is weakest on the sealed yes/no tasks, with competence between +8 and +19. I set out to test whether the same recipe at 4B, with more task variety, closes that gap.

Three findings mattered. A 4B base trained from raw beat the same recipe at 2B by +16.8 public points in a short pilot. New held-out task families raised scores on unseen tasks in a 2B screen, so I carried them to 4B. A full second stage on the post-trained 4B base, one full epoch instead of 1,500 steps, lifted the unseen-task score from 15.3 to 24.7.

The training recipe is the v20 recipe from the first article: a 27B teacher supplies yes/no targets only where it agrees with the gold label, with no frozen-base anchor on yes/no.

Unseen-v2 is a new set of 9,000 rows from task families no model trained on, split 40 percent dev and 60 percent test. It is harder than the first article's 1,050-row set, so scores are lower. Image results use NaturalBench (1,200 questions in 300 groups), POPE-adversarial (600) and 60 Image JevBench preview items.

Each pilot arm trained 1,500 steps from raw on the v20 recipe. Before any score, I fixed a gate: the better 4B arm had to beat the 2B by at least +5 public points and be no worse on unseen dev.

The gate passed at +16.8. The rule picks the higher public arm, Base, by 0.6 points, within noise. The post-trained arm generalised better on unseen dev by 2.5 points, and its checkpoint became the starting point for v22.

In parallel I screened three recipe changes on 2B in eight cells, every on and off combination, each continuing hobson-v20 for 1,500 steps. Each arm was judged on its own preregistered unseen-v2 dev metric.

Alt text: Main effects with 95 percent intervals: A1 new task families +3.83 intelligence (adopt), A2 RPS plus NLL +0.06 score competence (no effect), A3 pooled 27B teacher -13.0 yes/no competence (harmful).

A1 added new task families to the training mix and raised unseen intelligence by +3.83 with no guard drops, so I adopted it. A2 added a ranked probability score and a negative log-likelihood term for score questions; its +0.06 interval spans zero, so I rejected it.

A3 is the honest negative result. Pooling the 27B teacher with the gold label on every row made the model hedge more on yes/no questions: competence fell by 13.0, unseen intelligence fell by 4.0 and public fell by 1.9. The agree-with-gold teacher from v20 stays.

The screen also showed a 2B limit. Through the image stage, the A1 cell's unseen-v2 score fell from 11.1 to between 3.9 and 5.5, while my first 4B run kept its score (14.7 before, 14.7 to 15.3 after). That pushed the next run to 4B.

My first 4B run had been short for budget reasons, 1,500 steps without A1. For v22 I continued the post-trained pilot checkpoint for one full epoch, 4,465 steps over 142,903 rows of the v20 mixture plus the A1 families, with two seeds. The interpolation check between the seeds showed no barrier (-0.004), so I averaged the weights into one soup. The text soup scored 62.4 public, got 191 of 231 tasks right, left 19 yes/no answers in the band and had a text ECE of 0.029, the lowest I have measured.

I then ran the hobson-v20-balanced image recipe on the soup with two seeds. A rule fixed before the run picked seed 0: the highest public score among seeds with images no worse than the published balanced model. The run took about 5.3 hours on two H200 GPUs.

Paired against the first 4B run on unseen-v2 test, v22 gains +9.36 intelligence (95% CI +7.92 to +10.83) and +25.2 yes/no competence (+21.8 to +28.6); choice (+1.3) is within noise. The public gain of the text soup is +6.3 with an interval from -0.1 to +13.1, so the clean evidence is on unseen tasks and yes/no.

Alt text: v19, hobson-v20-balanced and v22 on NaturalBench accuracy (0.787, 0.802, 0.818), group accuracy (0.347, 0.360, 0.400), POPE (0.858, 0.865, 0.875) and preview items (40, 43, 48 of 60).

Across both image seeds, v22 beats the published balanced model on NaturalBench group accuracy by +0.063 (+0.017 to +0.113) and on preview items by +7.5 (+1.0 to +14.0); POPE (+0.011) is inside noise. With the image removed it is 0.052 less confident on POPE than the balanced model, a sign it leans on the picture more. Seed 1 had the better image numbers, but the rule picks on public text.

For the sealed half I convert unseen-v2 to its equivalent on the older 1,050-row set, about 50, and add the offset I measured on decider-2b and decision-2b (+7.8 to +14.9), giving about 61. Averaged with the open half (64.6), Intelligence is about 63 (58.6 to 66.6). I assume Calibration near 87 (78 to 90), because the ranked Qwen3.5-4B deciders score 85.6 to 88.6, so Capability is about 75 (68 to 78).

Alt text: Estimated Capability band for v22, 68 to 78 with a middle of about 75, against jevk5-v0.3-4b 72.3, plumb-4b 71.6, decision-4b-v12 71.2 and decider-4b-v2 70.7.

At the middle estimate v22 is #1 in the 4B class, ahead of jevk5-v0.3-4b at 72.3, and around #5 to #7 at the pessimistic end. The overall board combines Capability with Speed and Cost; at decider-4b-v2's Speed and Cost, v22 maps to about 74, against 73.70 for Cygnet and 73.23 for Winnow-12B. Gaps under about 0.5 do not separate there, so my estimate is #1 to #2 overall, tied with Cygnet, and about #5 pessimistic. On the image board the 4B to beat is Imajev-4B, #3 with Capability 82.1; my range for v22 is roughly #3 to #15, because 60 preview items cannot place it more precisely. Every rank here is a local, unofficial estimate pending an official run.

Every text score comes from my reimplementation of the v1.5 rules, and the sealed half and Calibration are estimated. The final text and image stages ran two seeds each. Unseen-v2 is my own set, and I have not checked that its ranking agrees with the board. The 2B factorial ran with fused kernels and the 4B pilot without them, so comparisons across those runs are loose.

The official run settles the ranks. In the issue below I ask the maintainers to review v22 and, if it holds up, submit it to Benchmark Heaven for text and Image JevBench. The rank check I planned as Stage C was not run, and checking unseen-v2 against the board ranking is open work.

The proposal and every comparison are in issue #39  on strands-labs/strands-decider. The code, tests, model doc and configs are on my branch pr/7-hobson-v22 , ready to open as Part 7 of my pull request series once a review slot frees, and strands-decider serve DIR --vision runs the checkpoint. The weights are at strands-decider-4B-hobson-v22.tar , SHA-256 7305d520e83e3dd0c112e22da2ded26713ed6f911e2ff071d82e484f8d55c029. No official submission has been made yet.