startlux-models/StartLux-Decision-v1-35B-A3B

TEXT GENERATIONPricing:Input $0.4 / Cached $0.07 / Output $4Concurrent Unit Cost:2Model Size:35.1BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 1, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

StartLux-Decision-v1-35B-A3B is a 35.1 billion parameter mixture-of-experts model developed by StartLux-models, with approximately 3 billion active parameters per token, designed for decision-making tasks. It processes typed questions about a given state (text, JSON, or images) up to a 262,144-token context, returning probabilities for each option. This model excels at providing fast, probabilistic answers for choices, yes/no questions, and ratings, supporting vision inputs and offering high performance on JevBench and Intern-Decision benchmarks.

Loading preview...

StartLux-Decision-v1-35B-A3B Overview

StartLux-Decision-v1-35B-A3B is a 35.1 billion parameter mixture-of-experts model from StartLux-models, optimized for rapid decision-making. It processes typed questions about a given state, which can be text, JSON, or images, and provides probabilistic answers for choices, yes/no questions, or ratings on a scale. With an effective 3 billion active parameters per token, it offers significantly faster inference compared to its 27B predecessor, answering in half the time.

Key Capabilities

  • Probabilistic Decision-Making: Provides a probability for every option in response to typed questions.
  • Multi-modal Input: Supports text, JSON, and image inputs, with a vision tower included for image processing.
  • Extended Context Length: Handles long inputs up to 262,144 tokens (256K), reading long states in chunks.
  • High Performance: Achieves 210 on JevBench public and 92.29% average accuracy on Intern-Decision, with a Decision Index 0.2.1 score of 61.55.
  • Fast Inference: Engineered for low latency, with a single yes/no question answered in 36.4 ms and three questions in 52.5 ms on an H200 GPU, utilizing fast kernels and CUDA graphs.
  • TypeSafe API: Uses the /v1/systemone format, ensuring compatibility with existing clients.

Good For

  • Automated Decision Systems: Ideal for applications requiring quick, probabilistic answers to structured questions.
  • Customer Support Automation: Can be used to triage tickets, assess urgency, or categorize issues based on textual and visual evidence.
  • Data Analysis: Extracting structured decisions or ratings from complex, multi-modal data.
  • Real-time Applications: Its low latency makes it suitable for scenarios demanding immediate responses.
  • Systems requiring vision-based decision support: Capable of incorporating image data into its decision process.

Finetunes
1 models