startlux-models/StartLux-Decision-v1-4B

Hugging Face
VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Sep 29, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Warm

StartLux-Decision-v1-4B is a 4.5 billion parameter model developed by startlux-models, designed for making probabilistic decisions from various input states. It excels at answering typed questions (choice, yes/no, rating) about text, JSON, or image inputs, providing a probability for each option. With a context length of up to 262,144 tokens and integrated vision capabilities, it is optimized for fast inference and high accuracy in decision-making tasks, particularly for structured queries.

Loading preview...

StartLux-Decision-v1-4B Overview

StartLux-Decision-v1-4B is a 4.5 billion parameter model from startlux-models, specifically engineered for probabilistic decision-making. It processes typed questions (e.g., multiple choice, yes/no, scaled ratings) against diverse input states, including text, JSON, and images, returning a probability distribution for potential answers. This model is part of a family of decision models, offering a balance of performance and speed.

Key Capabilities

  • Probabilistic Decision-Making: Provides a probability for each option in response to structured questions.
  • Multimodal Input: Supports text, JSON, and image inputs, with a built-in vision tower for image processing.
  • Extended Context Window: Handles long inputs up to 262,144 tokens (256K), allowing for comprehensive state analysis.
  • High Accuracy: Achieves strong results on decision-making benchmarks, including 204 correct answers on JevBench public and 91.17% average accuracy on Intern-Decision.
  • Fast Inference: Optimized for low-latency inference, with a single request involving three questions completing in 26.0 ms on an H200 GPU, utilizing fast kernels and CUDA graphs.
  • Batch Processing: Features decide_batch for efficient processing of multiple requests.
  • TypeSafe API: Uses the /v1/systemone format, ensuring compatibility with existing Jev clients.

Good For

  • Automated Triage and Routing: Categorizing customer support tickets or other inquiries based on content and urgency.
  • Structured Data Analysis: Extracting specific decisions or ratings from complex text or JSON data.
  • Image-Based Decision Support: Answering questions about visual content, such as identifying damage in photos.
  • High-Throughput Decision Systems: Applications requiring rapid, probabilistic answers to structured queries over large datasets.

Finetunes
4 models