PNPM244/linne-oak-1-0

VISIONPricing:Input $0.48 / Cached $0.1 / Output $1.44Concurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 11, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

PNPM244/linne-oak-1-0 is a 31 billion parameter decision model developed by Minh Duc Tran, fine-tuned from google/gemma-4-31B-it. It is designed to take a state (text, document, JSON, conversation, or agent trajectory) and a question with options, returning a single option with probabilities in one forward pass. This model excels at decision-making tasks, achieving a tied #1 ranking on the Decision Index 0.3.2, and is optimized for agent safety and fast inference with a 32768 token context length.

Loading preview...

linne-oak-1.0: A Specialized 31B Decision Model

PNPM244/linne-oak-1-0 is a 31 billion parameter model developed by Minh Duc Tran, specifically engineered for decision-making tasks. Unlike general-purpose causal language models, linne-oak-1.0 processes a given state (text, JSON, conversation, etc.) and a question with options to return a single, probabilistic decision in one forward pass. It is fine-tuned from google/gemma-4-31B-it on a diverse dataset covering classification, NLI, retrieval, tool selection, and agent safety.

Key Capabilities

  • Top-tier Decision Performance: Achieved a tied #1 ranking on the Decision Index 0.3.2 leaderboard (as of 2026-10-14), demonstrating strong performance across public benchmarks and private tasks, including new domains.
  • Full Prompt Attention: Utilizes full-attention layers that attend bidirectionally over the entire prompt, enabling comprehensive understanding of complex inputs up to 32768 tokens.
  • Agent-Safe: Designed with a strong step-level guard for agents, featuring a low false-block rate, making it suitable for integration into autonomous systems.
  • Fast Inference: Delivers rapid decision outputs, with a median inference time of 105 ms per request on an RTX PRO 6000 and 63 ms on an H200.
  • Broad Task Coverage: Strong performance across knowledge, language understanding, retrieval, tool use, and automation tasks.

Good For

  • Automated Decision Systems: Ideal for applications requiring a model to select from a set of options based on complex input states, such as routing customer support requests or automating agent actions.
  • Agentic Workflows: Provides a robust and safe decision-making component for AI agents, ensuring reliable and guarded step-level choices.
  • Classification and Retrieval: Excels in tasks that involve classifying inputs or retrieving the most relevant option from a given set.
  • High-Performance Inference: Suitable for scenarios where fast, single-pass decision outputs are critical, leveraging its optimized architecture for speed.

Finetunes
1 models