KnowLine-9B-Gen2

The 9B Gen2 release of PelaAI's System One decision model: 9B parameters, trained, like the other KnowLine models, through a data iteration and research loop driven by the same agent on an all-in-one computer. Served with read twice.

Inference guide ยท ไธญๆ–‡่ฏดๆ˜Ž ยท weights Apache-2.0 ยท siblings: KnowLine-26B-A4B-Gen2.5, KnowLine-35B-A3B-Gen1.5 ยท previous release: KnowLine-9B-Gen1

Highlights:

  • Decision Index 0.3 public suite: 69.00 (self-run). KnowLine-35B-A3B-Gen1.5 and KnowLine-26B-A4B-Gen2.5 score 70.72 and 68.21 on the same suite with the same serving, respectively; see Comparison.
  • Calibrated probabilities: a temperature, fitted on held-out data, is folded into the weights. Computed with the board's method on the public suite, ECE is about 0.009โ€“0.011 and the Brier score about 0.253โ€“0.272 (KnowLine-26B-A4B-Gen2.5: 0.013-0.018 and 0.24-0.27); see Calibration.

KnowLine is an independently developed model. It is not affiliated with or endorsed by TypeSafe, Jev or Qwen, and it is not derived from TypeSafe or Jev.

What it does

You send a state (text or a conversation) 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.
Base model Qwen/Qwen3.5-9B (Apache-2.0): dense, about 9B parameters
Training LoRA supervised fine-tuning of the language model component, merged into bf16 weights (see How Gen2 was made)
Release format bf16 weights (about 19.3 GB), with a readout temperature T = 1.63 folded into the final norm weight (see Calibration); the front end reads score questions at an effective temperature of 1.0 (temperatures.json). The vision component is unchanged from the base model; the tokenizer is Qwen3.5's.
Serving serve_knowline.sh: SGLang in bf16, as for the 26B-A4B; knowline_server.py reads short requests twice, as in KnowLine-26B-A4B-Gen2.5 and KnowLine-35B-A3B-Gen1.5 (on by default; --no-read-twice reads each request once). SGLang reserves 85% of the GPU's memory by default (MEM); the weights take about 19.3 GB. This DI evaluation used H20 96 GB hardware.
Languages English, Simplified Chinese, Traditional Chinese

Quickstart

pip install "sglang==0.5.21" "transformers==5.12.1" requests
bash serve_knowline.sh PelaAI/KnowLine-9B-Gen2 0 8080          # SGLang (bf16) on :9080, /v1/systemone on :8080
curl -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 with knowline_layout.py next to it; together they depend only on transformers and requests. Text requests are rendered with Qwen3.5's chat template.
  • Full settings: the exact settings used for our Decision Index run are in INFERENCE.md.

How Gen2 was made

  • Base model: Qwen/Qwen3.5-9B (Apache-2.0), a dense model.
  • Training: the language model component was trained with a LoRA adapter (rank 128, alpha 256, language model only, one epoch), merged into bf16 weights. Training examples include decision tasks written and labelled by LLM teachers, rendered with Qwen3.5's chat template.
  • Serving and temperature: it is served with read twice, as KnowLine-26B-A4B-Gen2.5 and KnowLine-35B-A3B-Gen1.5. A single readout temperature T = 1.63, fitted on held-out data with this serving, is folded into the weights. It changes the size of the probabilities, but not which option ranks first.

The training data is not included in this repository.

Comparison

The table below shows Decision Index scores. Since Decision Index 0.4, the board's full score is 60% private tests of the same skills and 40% private tasks from new domains; the public suite is shown but no longer counts toward the full score. This model has no board entry yet.

Model Size DI 0.3 public DI 0.4.1 full Source
KnowLine-9B-Gen2 9B 69.00 Not yet scored Self-run, official kit
karpo-decision-31B About 31.3B 75.05 73.86 Board
GLiDE 27B About 26.1B 72.49 73.37 Board
REAX 1.3 About 57.4B 73.71 72.64 Board
Logits Sigma-1 Duo About 57.4B 72.93 72.28 Board
Copernicus-31B About 31.3B 70.37 72.11 Board
KnowLine 31B Gen1.5 About 31.3B 70.50 71.91 Board
Copernicus-57B About 57.4B 75.34 71.80 Board
FlexDM-LoRA 2.1 ยท Gemma-4-31B About 31.3B 68.99 71.61 Board

Board rows are from the official Decision Index data, updated on 2026-11-12.

Public-suite skill by area (DI 0.3, all three served with read twice):

Area KnowLine-9B-Gen2 KnowLine-26B-A4B-Gen2.5 KnowLine-35B-A3B-Gen1.5
Knowledge & reasoning 52.3 55.1 55.0
Language 69.5 71.2 72.9
Retrieval & routing 75.1 70.9 75.4
Tools & agents 88.8 83.3 88.2
Arts & taste 62.6 61.4 64.2

Releases

The model is trained in a self-evolving loop, and new versions will follow.

Model Version date DI 0.3 public Notes
KnowLine-31B-Gen2 2026-11-13 73.00 Second 31B release
KnowLine-12B-Gen1 2026-11-13 69.37 First 12B release
KnowLine-9B-Gen2 (this model) 2026-11-13 69.00 Second 9B release
KnowLine-31B-Gen1.5 2026-11-04 70.50 Gen1 weights; read twice, with a question list for long requests
KnowLine-2B-Gen1 2026-10-29 52.23 First 2B release
KnowLine-0.8B-Gen1 2026-10-29 44.51 First 0.8B release
KnowLine-9B-Gen1 2026-10-25 65.67 First 9B release
KnowLine-31B-Gen1 2026-10-24 70.13 First 31B release
KnowLine-26B-A4B-Gen2.5 2026-10-22 68.21 Gen2 weights, read-twice serving
KnowLine-35B-A3B-Gen1.5 2026-10-22 70.72 35B-A3B Gen1 weights, read-twice serving
KnowLine-4B-Gen5.5 2026-10-22 66.75 Gen5 weights, read-twice serving
KnowLine-4B-Gen5 2026-10-21 65.67 4B, fifth release
KnowLine-26B-A4B-Gen2 2026-10-19 66.94 26B-A4B, second release
KnowLine-26B-A4B-Gen1 2026-10-19 65.76 First release on a Gemma 4 base
KnowLine-35B-A3B-Gen1 2026-10-12 69.49 First 35B-A3B release
KnowLine-4B-Gen4 2026-10-10 64.90 4B, fourth release
KnowLine-4B-Gen3 2026-10-09 63.11 4B, third release
KnowLine-4B-Gen2 2026-10-08 62.54 4B, second release
KnowLine-4B-Gen1 2026-10-07 60.47 4B, first release

Evaluation

All results below are self-run and have not been verified by a third party. Every number for this model was measured through its release front end (read twice on). The 26B-A4B-Gen2.5 and 35B-A3B-Gen1.5 numbers are from their release evaluations with the same front end.

Images (zero-shot)

  • No image training: the vision tower is the base model's, unchanged, and no images were seen 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. Requests with images are read once (read twice is text-only).

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.

Metric 9B-Gen2 26B-A4B-Gen2.5 35B-A3B-Gen1.5
ECE (lower is better) 0.009โ€“0.011 0.013-0.018 0.03-0.04
Brier score (lower is better) 0.253โ€“0.272 0.24-0.27 0.24-0.26
Confidence โ‰ฅ95% but wrong 0.9% 0.9% 1.5%
Mean confidence / accuracy 0.81โ€“0.82 / 0.80โ€“0.82 0.82-0.83 / 0.80-0.82 0.85-0.86 / 0.81-0.83
  • Temperature in the weights: the final norm weight includes a readout temperature T = 1.63, fitted on held-out data with the read-twice serving, so every final logit is divided by 1.63 before the softmax over the option labels. The top-ranked option does not change.
  • Score questions at temperature 1.0: serve_knowline.sh passes temperatures.json ({"score": 0.6135}), so the front end divides the logits of score (rubric) questions by a further 0.6135 and reads them at an effective temperature of 1.63 ร— 0.6135 = 1.0; yes/no and choice questions stay at 1.63. Score questions are often read as an expected value over the scale, which a higher temperature pulls toward the middle. This is a serving setting, not a weight change, and earlier KnowLine releases have used it; starting knowline_server.py without --temperatures (or knowline_engine.py with --option temperatures=none) reads every question at 1.63.
  • Why ranges: the board does not publish which benchmarks it uses for calibration, so we report ranges over the plausible benchmark sets, as for the earlier releases.
  • Recommendation: if your use depends on probability thresholds, check calibration on your own data.

Disclosures

  • Decontamination: standard and strict decontamination checks were performed when assembling the training data for this run, removing 63,444 rows during assembly.
  • 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.
  • Public suite: the Decision Index public suite is public; the board's private tests may show a lower score.

Limitations

  • Size and speed: about 19.3 GB of bf16 weights and dense, so every token uses all 9B parameters; this serving used a 96 GB H20. One request at a time on one H20, requests take 97.9 ms median and 209.4 ms mean (833.7 ms at the 95th percentile; 300 requests in this independent latency run). The historical measurements for KnowLine-26B-A4B-Gen2.5 are 66.3 ms and 151 ms; see INFERENCE.md.
  • Read twice: text-only requests up to 3,000 tokens are written twice. Use --no-read-twice if latency matters more; we did not evaluate this model without read twice for this card.
  • Planted instructions: this model can be steered by instructions planted in the input; put untrusted content in data fields of the state, not in the instructions, and keep it clearly delimited.
  • Calibration: probabilities are calibrated on average, not for every domain; check calibration on your own data.
  • Images: there was no image training, so check image use on your own data.
  • Other attacks: other prompt-injection attacks may still work.
  • Private tests: the board's private tests may show a lower score than the public suite.

Citation

@misc{knowline9bgen2,
  title  = {KnowLine-9B-Gen2: a 9B decision model},
  author = {PelaAI},
  year   = {2026},
  url    = {https://huggingface.co/PelaAI/KnowLine-9B-Gen2}
}

License

  • Weights: Apache-2.0, the same as the base model.
  • Base model licence: Qwen/Qwen3.5-9B is released under Apache-2.0; see the Qwen3.5-9B licence file.
  • Code: knowline_server.py, knowline_layout.py and knowline_engine.py are MIT (see the file headers).
Downloads last month
25
Safetensors
Model size
10B params
Tensor type
F32
ยท
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for PelaAI/KnowLine-9B-Gen2

Finetuned
Qwen/Qwen3.5-9B
Finetuned
(1,389)
this model