torchcast-ai/torchcast-decision-12b

TEXT GENERATIONPricing:Input $1.2 / Cached $0.24 / Output $4.8Concurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Oct 2, 2026License:cc-by-nc-4.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Torchcast AI's Torchcast Decision 12B is a 12-billion parameter text-based decision checkpoint, fine-tuned from Gemma-4-12B-it, designed for structured decision-making tasks. It supports yes/no probabilities, choice distributions, and ordinal score distributions, with a served context limit of 16,384 tokens. This model excels at classification, routing, and bounded rubric evaluation, providing typed decisions via a specialized serving package rather than generic text generation.

Loading preview...

Torchcast Decision 12B: Specialized Decision-Making Model

Torchcast Decision 12B, developed by Torchcast AI, is a 12-billion parameter model based on Gemma-4-12B-it, specifically engineered for text-based decision checkpointing. Unlike general-purpose LLMs, this model is optimized to output structured decisions, including yes/no probabilities (noul), choice distributions (choice), and ordinal score distributions (score). It operates with a served context limit of 16,384 tokens.

Key Capabilities

  • Structured Decision Outputs: Provides probabilistic and distributional outputs for clear, actionable decisions.
  • Specialized Fine-tuning: LoRA fine-tuned on a mix of gold labels and instruction model distributions, including procedurally generated decisions and public datasets.
  • Evaluated Interface: Designed to be used with a supplied serving package for typed decisions, ensuring evaluated interfaces are reproduced accurately.
  • Performance: Achieved 87.88% correctness (203/231) on the JevBench v1.4 public set (author-run, not official), with low latency for easy and original items.

Good For

  • Classification and Routing: Ideal for tasks requiring precise categorization or directing inputs based on textual analysis.
  • Bounded Rubric Evaluation: Suitable for automated assessment against defined criteria.
  • Decision Research: Supports research into text-based decision processes, particularly where probabilistic or distributional outcomes are needed.
  • Non-Commercial Use: Licensed under CC BY-NC 4.0, making it suitable for academic and research projects.

Finetunes
1 models