SkyPanther/synack-decide-26b-a4b

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 4, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SynACK/synack-decide-26b-a4b is a 26 billion parameter open "System One" decision model, fine-tuned from google/gemma-4-26B-A4B-it. It is designed to provide single-pass, probabilistic answers to multiple-choice (2-255 options) or yes/no questions based on a given JSON state, without generating reasoning text. This model excels at rapid decision-making tasks, achieving a Decision Index 0.2.1 score of 59.5, and is particularly strong in areas like PR review routing and handling questions with many options.

Loading preview...

SynACK Decide 26B-A4B: A Specialized Decision Model

SynACK Decide 26B-A4B is an open "System One" decision model, fine-tuned from google/gemma-4-26B-A4B-it. This 26 billion parameter model (with approximately 4 billion active per token) is engineered for rapid, single-pass decision-making, providing probabilistic answers to multiple-choice or yes/no questions based on a JSON state input. Unlike generative LLMs, it produces no reasoning text, focusing solely on efficient, direct answers.

Key Capabilities

  • Direct Decision-Making: Returns one answer per question with probabilities for each option from a single forward pass, enabling decisions in hundreds of milliseconds.
  • High Performance on Decision Index: Achieved a Decision Index 0.2.1 score of 59.5, outperforming several reference models on the public leaderboard.
  • Specialized for PR Review Routing: Demonstrates an AUC of 0.774 for predicting if a pull request needs substantive changes, surpassing a deterministic complexity script.
  • Handles Multi-Option Questions: Capable of answering questions with up to 255 options, addressing a limitation of smaller models.
  • Flexible Deployment: Supports transformers (CUDA), Apple Silicon (MLX), and llama.cpp (GGUF) for diverse environments.

Good For

  • Automated Triage and Routing: Ideal for systems requiring quick, probabilistic decisions, such as routing customer support tickets or pull requests.
  • Structured Question Answering: Excels in scenarios where questions have predefined options, like surveys, classifications, or diagnostic tools.
  • Low-Latency Applications: Its single-pass nature makes it suitable for applications where speed is critical and generative text is not required.
  • Integrating with Existing Systems: Speaks the Jev / Decision Index /v1/systemone wire format, facilitating integration into existing decision-making pipelines.

Finetunes
3 models