torchcast-ai/torchcast-decision-27b

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 3, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Torchcast-AI's Torchcast Decision 27B is a 27 billion parameter language model fine-tuned for decision-making tasks, specifically answering typed questions about a state with probabilistic outputs for choices, yes/no questions, or ordinal ratings. It achieves a Decision Index 0.2.1 score of 65.00, outperforming Jev 1.13, and supports a context length of up to 262,144 tokens. This model is optimized for fast inference, processing a single yes/no question in 33ms on an H100 GPU, making it suitable for applications requiring rapid, probabilistic decision outputs.

Loading preview...

Torchcast Decision 27B Overview

Torchcast Decision 27B, developed by Torchcast-AI, is a 27 billion parameter language model specifically designed for probabilistic decision-making. It is a LoRA fine-tune of StartLux-Decision-27B, merged into the weights, and excels at answering typed questions about a given state, providing probabilities for each option in choice, noul (yes/no), and score (ordinal rating) question types. The model supports a substantial context length of up to 262,144 tokens.

Key Capabilities

  • Probabilistic Decision Outputs: Provides a probability for every option in response to typed questions, enabling nuanced decision support.
  • High Performance: Achieves a Decision Index 0.2.1 score of 65.00, surpassing models like StartLux-Decision-27B (63.88) and Jev 1.13 (57.91).
  • Fast Inference: Optimized for speed, capable of answering a single yes/no question in 33.0 ms and a three-question request in 49.0 ms on an H100 GPU, utilizing fast kernels and CUDA graphs.
  • Batch Processing: Supports decide_batch for efficient processing of multiple requests.
  • Broad Task Coverage: Shows significant gains over Jev 1.13 in areas like tool use (home appliance simulator), aspect-sentiment extraction, stance detection, and grade-school math.

Good For

  • Automated Decision Support: Ideal for systems requiring rapid, probabilistic answers to structured questions based on a given state.
  • Research and Non-Commercial Use: Intended for academic and non-profit applications, particularly in areas requiring high-throughput decision analysis.
  • Applications with High Context Needs: Suitable for scenarios where extensive textual context (up to 256K tokens) is necessary for decision-making.

Limitations

  • Knowledge Area: Trails Jev 1.13 in hard knowledge benchmarks (e.g., GPQA Diamond, MMLU-Pro, BBH).
  • English Only Evaluation: While the base model supports image inputs, evaluations for this fine-tune were text-only and in English.
  • Non-Commercial License: The model weights are licensed under CC BY-NC 4.0, restricting commercial use without separate permissions.

Finetunes
1 models