oak1.1

oak1.1 is a 31B decision model. Give it a state (text, a document, JSON, a conversation or an agent trajectory) and a question with a set of options, and it returns one option together with a probability for every option, in a single forward pass.

It is the successor of linne-oak-1.0, with the same interface and serving code.

Highlights

• 68.19 on the Decision Index public suite (complete run, 140,620 requests): +4.25 over linne-oak-1.0 (63.94).
• Better in every area: arts +12.4, tools +7.2, language +4.3, retrieval +2.7, knowledge +0.1.
• Sees all the questions first: when a request asks several questions about the same state, the prompt lists them before the state, so the state is read knowing what will be asked.
• Reads the whole prompt at once: the full-attention layers attend in both directions over the prompt (sliding-window layers keep their native behavior).
• Fast: one forward pass per decision; median 67 ms per request on one H200.

Decision Index, public suite

Complete run of the public 0.3 suite (140,620 requests, all answered), scored with the leaderboard's code. Results: PNPM244/oak1.1-decision-index. Official Full scores (private tests run by the maintainers) will be added when available.

Model	Public index	Knowledge	Language	Retrieval	Tools	Arts
oak1.1 (31B)	68.19	54.4	70.9	71.8	85.7	57.5
linne-oak-1.0 (31B)	63.94	54.3	66.6	69.1	78.5	45.1

Training data

Fine-tuned from google/gemma-4-31B-it on decision-format data (state, question, options) covering classification, NLI, retrieval, tool selection and agent safety. The training data is not released.

Usage

The prompt is one user message (chat template, thinking off): a short list of the questions asked about the state, Context: and the state (compact JSON when not a string), a blank line, Question: <instructions>, one line <label>: <name> - <description> per option, and Answer with the letter of one option only. The answer is the softmax over the next-token logits of the option letters (with the model's final logit soft-capping), read at the last prompt position, with the full-attention layers bidirectional.

linne_engine.py (with nc_hook.py) in this repository implements exactly this, as a Decision Index engine and as a library. Its defaults (noncausal=True, layout="header") are the settings this model was evaluated with.

import sys
from huggingface_hub import snapshot_download

path = snapshot_download("PNPM244/oak1.1")
sys.path.insert(0, path)
from linne_engine import LinneEngine

model = LinneEngine(model=path, device="cuda:0")
result, _ = model(
    "My Stripe integration keeps failing. Please help ASAP.",
    {"team": {"type": "choice", "instructions": "Which team should handle this request?",
              "criteria": {"billing": "Charges and refunds", "technical_support": "Integration errors",
                           "sales": "Questions about buying a product"}}},
)
print(result["answers"]["team"])  # chosen option and a probability for every option

Run the Decision Index public suite:

PYTHONPATH=<path> python -m decision_index pipeline --engine linne_engine:LinneEngine \
  --option model=<path> --option device=cuda:0 --option noncausal=1 --option layout=header --out runs/oak1.1

BF16 weights are about 62 GB and fit one 80 GB GPU.

License

CC BY-NC 4.0: free for non-commercial use with attribution. Commercial use requires prior written permission from the author.

Built on Gemma 4 (Apache 2.0); see NOTICE and LICENSE-APACHE-2.0.

Citation

@misc{oak_1_1,
  title        = {oak1.1: A 31B Non-Causal Decision Model},
  author       = {Tran, Minh Duc},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/PNPM244/oak1.1}},
  note         = {Version 1.1}
}