j-raghavan/s1-gemma4-26b-decision

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:2Model Size:26BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 7, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

j-raghavan/s1-gemma4-26b-decision is a 26 billion parameter mixture-of-experts model, fine-tuned from google/gemma-4-26B-A4B-it, designed for calibrated decision-making. It processes a given state and question to return probabilities for predefined options in a single forward pass, without generating text. This model excels at structured decision tasks, offering specific question types like 'choice', 'noul' (yes/no), and 'score', making it suitable for applications requiring probabilistic, explainable decisions rather than generative text.

Loading preview...

s1-gemma4-26b-decision: A Calibrated Decision Model

s1-gemma4-26b-decision is an open, calibrated "System One" decision model, fine-tuned from Google's Gemma 4 26B-A4B-it. Unlike traditional LLMs that generate text, this model is specifically designed to provide probabilistic answers to structured questions in a single forward pass. It takes a state (text or JSON) and a typed question, then returns a probability for every option, making it ideal for automated decision systems.

Key Capabilities

  • Probabilistic Decision Making: Provides a probability for each option in choice, noul (yes/no), and score question types.
  • Efficient Inference: Returns results in one forward pass, with no text generation, enabling low-latency decisions (median 27 ms in FP8).
  • Calibrated Outputs: Offers calibrated probabilities, improving reliability for critical applications.
  • Flexible Deployment: Supports deployment via vLLM for GPU servers and Ollama for local, quantized inference (int4).
  • Robust Training: Fine-tuned on over 100,000 typed decision rows from diverse public datasets, including intent, NLI, QA, and prompt injection sets.

Good For

  • Automated Routing and Triage: Directing customer support tickets, classifying incidents, or routing requests based on probabilistic assessments.
  • Risk Assessment: Evaluating conditions and providing a probability of 'yes' for specific risks.
  • Scoring and Ranking: Assigning probabilities to ordered levels, such as severity scores or confidence levels.
  • Applications Requiring Explainable Decisions: When a clear, probabilistic outcome is needed rather than a generated explanation, allowing for human oversight on uncertain cases.
  • Resource-Constrained Environments: The int4 quantized version allows for local deployment on consumer hardware like Apple-silicon laptops.

Finetunes
1 models