What it does
You send a state (text or a chat) and up to 64 typed questions: yes/no, choose one of k, or score on a rubric. The model
answers every question with a probability distribution over its options:
- one forward pass per question, with no text generation;
- the output is the probability of each option's label token;
- existing Jev clients only need to change the base URL, and users of other KnowLine models only need to change the model
name, because the interface is the same.
Table | |
|---|
| Base model | google/gemma-4-26B-A4B-it (Apache-2.0): mixture of experts, about 26B parameters in total and about 4B active per token |
| Training | LoRA SFT (rank 32, alpha 64, language model only, including the routed experts), merged into bf16 weights; then a short second LoRA round, also merged |
| Release format | bf16 weights (about 52 GB), with a readout temperature T = 1.25 folded into the final norm weight and the final logit soft-cap (see Calibration). The vision tower is unchanged from the base model; the tokenizer is Gemma 4's, not the one of the Qwen-based KnowLine models. |
| Serving | serve_knowline.sh: SGLang in bf16. We do not convert the weights to FP8 at load for this model: in our tests two FP8 runs of the same questions agreed on only 95.7% of the decisions, against 99.4% for bf16. SGLang reserves 85% of the GPU's memory by default (MEM; about 82-88 GB on a 96 GB card); the weights take about 52 GB, so one 80 GB GPU is enough. The transformers backend peaks at about 60 GB. We ran every evaluation with one server on one 96 GB H20. |
| Languages | English, Simplified Chinese, Traditional Chinese |
Quickstart
pip install "sglang==0.5.21" "transformers==5.12.1" requestsbash serve_knowline.sh PelaAI/KnowLine-26B-A4B-Gen1 0 8080 # SGLang (bf16) on :9080 + /v1/systemone on :8080curl -s http://127.0.0.1:8080/v1/systemone -H 'Content-Type: application/json' -d '{ "model": "m", "state": "Customer: my order arrived broken, I want my money back.", "questions": {"refund": {"type": "noul", "instructions": "Should the agent offer a refund?"}, "tone": {"type": "choice", "instructions": "Customer tone?", "criteria": {"angry": "Angry", "neutral": "Neutral", "happy": "Happy"}}}}'
- Front end:
knowline_server.py is a single file that depends only on transformers and requests. It is the same
code as in the other KnowLine releases; it renders each question with the model's own chat template.
- Full settings: the exact settings used for our Decision Index run are in INFERENCE.md.
How Gen1 was made
-
Why a Gemma base? KnowLine-35B-A3B-Gen1 is strong on the public suite and on private tests of the same skills,
but on decision tasks from new domains it trails models of similar size. On our own new-domain tasks, the stock
gemma-4-26B-A4B-it already reads such tasks better than the stock Qwen3.5-35B-A3B. The agent trained the latest data
round of the 35B line on this base to see how much of the new-domain gap comes from the base model.
-
Data. Two rounds:
- the latest data round of the 35B line, rendered with Gemma's chat template;
- a short second round of injection-invariance items (instructions planted in the state that must not change the
answer), plus a replay share of the first round.
All data went through the same decontamination check as the earlier releases; see Disclosures.
-
Training. One epoch on the first round, then one epoch on the second round, each with a LoRA adapter merged into
the weights. A single temperature was fitted and folded into the weights. It changes the size of the probabilities,
but not which option ranks first.
-
Selection. The first-round model passed our checks against the stock gemma-4-26B-A4B-it except one: planted instructions changed 6.6% of its answers,
above our limit of 4%. The second round brought this to 0.3% (unseen templates: 16.2% → 1.8%) without a significant
change on our held-out sets or Open-Jev 1.1 OOD, within our margin on new-domain tasks (10-task set 56.4 → 55.2;
34-task set 59.3 → 59.9), and with a significant gain on JevBench-hard (73.9 → 80.2), so it is the release. Against KnowLine-35B-A3B-Gen1, it is level on our English in-house set and
higher on JevBench-hard, but significantly lower on our Chinese in-house sets and on Open-Jev 1.1 OOD.
Comparison
Scores on the Decision Index 0.3 board. The full score also includes private tests that only the maintainers run
(weighted 0.5 same-skill and 0.3 new-domain); the full score of this model is not available yet.
Table with columns: model, size, DI 0.3 public, DI 0.3 full, source| model | size | DI 0.3 public | DI 0.3 full | source |
|---|
| KnowLine-26B-A4B-Gen1 | 26B-A4B | 65.76 | not yet scored | self-run, official kit |
| KnowLine-35B-A3B-Gen1 | 35B-A3B | 69.49 | 62.92 | board |
| StartLux-Decision v1.1 27B | 27B | 72.88 | 69.12 | board |
Board rows as of the board's latest update (2026-10-16).
Public-suite scores by area (DI 0.3):
Table with columns: area, KnowLine-26B-A4B-Gen1, KnowLine-35B-A3B-Gen1| area | KnowLine-26B-A4B-Gen1 | KnowLine-35B-A3B-Gen1 |
|---|
| Knowledge & reasoning | 50.4 | 52.5 |
| Language | 69.2 | 71.2 |
| Retrieval & routing | 69.7 | 75.2 |
| Tools & agents | 82.2 | 88.1 |
| Arts & taste | 58.6 | 63.4 |
Releases
The model is trained in a self-evolving loop, and new versions will follow.
Table with columns: model, date, DI 0.3 public, golden held-out (en / zh-Hans / zh-Hant), notes| model | date | DI 0.3 public | golden held-out (en / zh-Hans / zh-Hant) | notes |
|---|
| KnowLine-26B-A4B-Gen1 (this model) | 2026-10-19 | 65.76 | 70.6 / 72.2 / 69.2 | first release on a Gemma 4 base |
| KnowLine-35B-A3B-Gen1 | 2026-10-12 | 69.49 | 70.3 / 73.8 / 72.1 | first 35B-A3B release; DI 0.3 full 62.92 |
| KnowLine-4B-Gen4 |
Evaluation
All results are self-run and have not been verified by a third party.
Held-out and Chinese evaluations
Table with columns: suite, 26B-A4B-Gen1, 35B-A3B-Gen1, stock gemma-4-26B-A4B-it| suite | 26B-A4B-Gen1 | 35B-A3B-Gen1 | stock gemma-4-26B-A4B-it |
|---|
| In-house evaluation set, English | 70.6 | 70.3 | 66.9 |
| In-house evaluation set, Simplified Chinese | 72.2 | 73.8 | 71.8 |
| In-house evaluation set, Traditional Chinese | 69.2 | 72.1 | 68.3 |
| C-Eval (4 categories, macro) | 74.3 | 87.6 | 67.5 |
In a paired sign test against 35B-A3B-Gen1, the drops on the Simplified and Traditional Chinese in-house sets
(p = 0.011 and p = 0.008) and on Open-Jev 1.1 OOD (p < 0.001) are significant, and so is the gain on JevBench-hard
(p = 0.013); the difference on the English in-house set is not (p = 0.60). Against the stock gemma-4-26B-A4B-it, the
training raised C-Eval, Open-Jev 1.1 and our in-house sets and cut the injection rate from 60.9% to 0.3%.
Web operation (Mind2Web official test splits, evaluation only)
Table with columns: split, 26B-A4B-Gen1 element selection, 26B-A4B-Gen1 operation (balanced), 35B-A3B-Gen1 element selection, 35B-A3B-Gen1 operation (balanced)| split | 26B-A4B-Gen1 element selection | 26B-A4B-Gen1 operation (balanced) | 35B-A3B-Gen1 element selection | 35B-A3B-Gen1 operation (balanced) |
|---|
| test_task (websites seen in training, new tasks) | 87.1 | 92.2 | 94.4 | 97.4 |
| test_website (new websites) | 87.0 | 93.8 | 92.4 | 97.5 |
| test_domain (new domains) | 87.0 | 93.9 | 93.1 |
- Lower than 35B-A3B-Gen1 on all three splits.
- The task is to pick the target element from a candidate set: the target plus up to 5 other elements sampled from the
page. This is easier than the original Mind2Web protocol, so do not compare these results with the Mind2Web
leaderboard.
Images (zero-shot)
- No image training: Gen1 keeps the base model's vision tower unchanged and saw no images in training.
knowline_server.py accepts OpenAI-style image / image_url content parts in a chat state and passes them to
SGLang; each question is still one prefill.
- How we measured it: the same reconstruction of the Decision Index 0.3.1 vision public suite as for the earlier
releases (11 benchmarks, chance-corrected skill and the board's weights), run through the release server, 0 failed
requests. We also ran the official vision public suite (12,210 rows, complete).
Table with columns: benchmark, 26B-A4B-Gen1 (our reconstruction), 26B-A4B-Gen1 (official vision suite), stock gemma-4-26B-A4B-it (our reconstruction, same front end), 35B-A3B-Gen1 (our reconstruction)| benchmark | 26B-A4B-Gen1 (our reconstruction) | 26B-A4B-Gen1 (official vision suite) | stock gemma-4-26B-A4B-it (our reconstruction, same front end) | 35B-A3B-Gen1 (our reconstruction) |
|---|
| CV-Bench | 72.1 | 71.9 | 68.3 | 75.8 |
| BLINK | 58.3 | 51.7 | 50.9 | 59.8 |
| RealWorldQA | 51.9 | 49.8 | 47.2 |
- Above the stock model, below the 35B-A3B: 5.2 points higher than the stock gemma-4-26B-A4B-it on our
reconstruction (charts, infographics, web screenshots), but 6.3 points lower than KnowLine-35B-A3B-Gen1, mostly on
RealWorldQA, web screenshots and exam pages; Hateful Memes and R-Bench-M are slightly below the stock model.
- Indicative only: the reconstruction is ours, not the board's rows; the official-suite column is a self-run of the
official kit.
Game harness (KOF '98)
- Setup: single-bout character-mirror matches, 18 games per pair, argmax actions; the same settings as for the other
KnowLine releases.
- Result: 15-2-1 against Jev, 17-1-0 against StartLux-Decision-35B-A3B and 8-9-1 against
KnowLine-35B-A3B-Gen1; 40-12-2 overall.
- How the games ended: against Jev, 5 of the 15 wins were by K.O. and 10 at time-out; against StartLux-Decision-35B-A3B,
14 of 17 by K.O.; against KnowLine-35B-A3B-Gen1, 5 of 8 wins and 8 of 9 losses by K.O.
- Play style: heavy kick 35%, special_2 32%, heavy punch 16%, special_1 7%, special_3 7%.
- Caveats: 18 games per pair give wide intervals; this is a measured result in this harness, not general
fighting-game skill.
Calibration
Computed on the 0.3 public suite using the method described by the Decision Index board: each field is scored right or
wrong, confidence is the probability of the chosen option, benchmarks are weighted equally, and 10 equal-width bins are
used.
Table with columns: metric, 26B-A4B-Gen1 (self-run), 35B-A3B-Gen1 (self-run), 35B-A3B-Gen1 (board), board median (173 models)| metric | 26B-A4B-Gen1 (self-run) | 35B-A3B-Gen1 (self-run) | 35B-A3B-Gen1 (board) | board median (173 models) |
|---|
| ECE (lower is better) | 0.02-0.03 | 0.04-0.05 | 0.048 | 0.078 |
| Brier score (lower is better) | 0.27-0.29 | 0.25-0.27 | 0.261 | 0.46 |
| Confidence ≥95% but wrong | 1.1-1.2% | 1.7% | 1.4% |
- Checked against the board: for KnowLine-35B-A3B-Gen1, the board's own figures (ECE 0.048, Brier 0.261) fall inside
the ranges our computation gave, so we use the same computation here.
- Temperature in the weights: the final norm weight and the final logit soft-cap include a readout temperature
T = 1.25, so every final logit is divided by 1.25 before the softmax over the option labels. The top-ranked option
does not change, so neither do the Decision Index scores; the probabilities are just less extreme. T was fitted on a
set of our own evaluation items whose distribution is covered by the training data, not on the held-out sets or on
Decision Index rows. On our held-out sets, ECE with the temperature in the weights is 0.037 (35B-A3B-Gen1: 0.025).
- Why ranges: the board does not publish which 32 benchmarks it uses, so we report ranges over the plausible
benchmark sets, as for the earlier releases.
- Recommendation: the temperature was fitted on our data. If your use depends on probability thresholds, check
calibration on your own data.
Disclosures
- Decision Index format training data: in the first round, about 36% of the rows target Decision Index benchmarks in
their request format or a reworded version of it: train splits of public datasets rewritten in that format (for
example WinoGrande XL, ACOS and GSM8K) and synthetic items written in the same format. Another 0.7% are copies of the
same items rendered in other formats. The second round is one third injection-invariance items and two thirds a replay
of the first round.
- Decontamination: all training data was checked with the same tool as the earlier releases against every Decision
Index 0.3 row, all of our evaluation sets and the other public benchmarks we report (Intern-Decision, S1MB, JevBench,
MMDM). In the parts built for this release, 491 rows that nearly duplicated an evaluation item, repeated its question
or claim, or contained its whole state were removed; the replayed parts come from data that had already passed the
same check.
- Selection on evaluations: we compared the first-round checkpoints and the second-round model on our release checks;
the release is the second-round model, the only one within our injection limit. It is significantly below
KnowLine-35B-A3B-Gen1 on our Chinese sets and Open-Jev 1.1 OOD (see Evaluation).
- No model outputs as labels: no output of Jev or any other decision model was used as a training label.
- Teacher-labelled synthetic data: LLM teachers wrote and labelled our synthetic tasks. The labels were filtered, but
not all of them were checked by a human.
Limitations
- Chinese and Open-Jev: clearly lower than KnowLine-35B-A3B-Gen1 on Chinese (C-Eval 74.3 against 87.6; our Chinese
in-house sets) and on Open-Jev 1.1. For Chinese-heavy use, prefer the 35B-A3B.
- Size and speed: about 52 GB of bf16 weights, so it needs one GPU with 80 GB or more; each request takes 44.8 ms
median, one request at a time on one H20 (see INFERENCE.md).
- Web operation and images: below KnowLine-35B-A3B-Gen1 on our Mind2Web and vision evaluations; there was no image
training, so check image use on your own data.
- Prompt injection: planted instructions changed 0.3% of answers on our injection set and 1.8% with unseen templates.
Other attacks may still work, so keep untrusted text clearly delimited.
- Private tests: part of the public-suite score comes from benchmarks whose tasks are in our training data, so the
0.3 private tests may show a lower score.
Citation
@misc{knowline26ba4bgen1, title = {KnowLine-26B-A4B-Gen1: a 26B-A4B decision model}, author = {PelaAI}, year = {2026}, url = {https://huggingface.co/PelaAI/KnowLine-26B-A4B-Gen1}}
License
- Weights: Apache-2.0, the same as the base model.
- Code:
knowline_server.py is MIT (see the file header).