startlux-models/StartLux-Decision-v1-27B

Hugging Face
VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 29, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

StartLux-Decision-v1-27B by startlux-models is a 27 billion parameter decision-making model designed to answer typed questions about a given state, providing probabilities for each option. It supports diverse input types including text, JSON, and images, with a substantial context length of up to 262,144 tokens. This model excels at tasks requiring probabilistic decision outputs like classification, yes/no questions, and rating scales, and is optimized for fast inference with specialized kernels and batch processing capabilities.

Loading preview...

StartLux-Decision-v1-27B Overview

StartLux-Decision-v1-27B is a 27 billion parameter model developed by startlux-models, specifically engineered for probabilistic decision-making. It processes typed questions about a given state (text, JSON, or images) and returns a probability for each possible option. The model supports an extensive context length of up to 262,144 tokens, allowing for comprehensive analysis of long inputs.

Key Capabilities

  • Probabilistic Decision Outputs: Provides a probability for each option in response to questions like 'pick one of several options', 'yes or no', or 'rating on a scale'.
  • Multi-modal Input: Accepts text, JSON, and images as input, with a vision tower integrated for image processing.
  • Extended Context Window: Handles inputs up to 262,144 tokens, enabling detailed analysis of large documents or complex states.
  • Fast Inference: Utilizes specialized kernels (flash-linear-attention, causal-conv1d) and CUDA graph replays for optimized, low-latency inference, achieving 102.3 ms for three questions and 50.7 ms for a single yes/no question on an H200 GPU.
  • Batch Processing: Features decide_batch for efficient processing of multiple requests simultaneously.
  • TypeSafe API Compatibility: Uses the /v1/systemone format, ensuring compatibility with clients designed for Jev.

Good For

  • Automated Triage and Classification: Assigning tickets to teams, categorizing customer inquiries, or determining urgency with probabilistic confidence.
  • Risk Assessment and Scoring: Evaluating severity, impact, or likelihood on a defined scale.
  • Image-based Decision Making: Answering questions based on visual evidence, such as identifying damage in a parcel image.
  • Applications Requiring Probabilistic Confidence: Use cases where not just a decision, but the confidence level of that decision, is crucial.

Finetunes
4 models