startlux-models/StartLux-Decision-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-35B-A3B is a 35.1 billion parameter mixture-of-experts model developed by StartLux Labs, with approximately 3 billion parameters active per token. It is designed for decision-making tasks, providing probabilities for options in response to typed questions about a given state, which can include text, JSON, or images. The model supports a large context length of 262,144 tokens and offers fast inference, answering in half the time of its 27B predecessor while achieving high scores on JevBench and Intern-Decision benchmarks.

Loading preview...

StartLux-Decision-35B-A3B Overview

StartLux-Decision-35B-A3B is a specialized mixture-of-experts (MoE) model from StartLux Labs, designed to provide probabilistic answers to decision-oriented questions. With 35.1 billion parameters, it activates about 3 billion parameters per token, enabling efficient processing. This model excels at tasks requiring a choice from options, yes/no answers, or ratings on a scale, providing a probability for each outcome.

Key Capabilities

  • Decision-Making Focus: Optimized for structured decision tasks, returning probabilities for each potential answer.
  • Multimodal Input: Processes state information from text, JSON, and images, supporting a rich understanding of context.
  • Extended Context Length: Handles inputs up to 262,144 tokens (256K), allowing for comprehensive contextual analysis.
  • High Performance: Achieves strong results on decision-specific benchmarks like JevBench (210/231 correct) and Intern-Decision (92.29% average accuracy).
  • Fast Inference: Engineered for speed, it answers in half the time of StartLux-Decision-27B, with a latency of 52.5 ms for three questions on an H200 GPU.
  • TypeSafe API: Uses the /v1/systemone format, ensuring compatibility with existing clients.

Good For

  • Applications requiring probabilistic decision outputs from complex inputs.
  • Systems needing to classify, rate, or make choices based on textual, structured, or visual data.
  • Use cases demanding high throughput and low latency for decision queries.
  • Integrating decision intelligence into existing systems using the TypeSafe API standard.

Finetunes
1 models